Báo cáo khoa học: "PROCESSING ENGLISH WITH A GENERALIZED PHRASE STRUCTURE GRAMMAR" pdf

PROCESSING ENGLISH WITH A GENERALIZED PHRASE STRUCTURE GRAMMAR Jean Mark Gawron, Jonathan King, John Lamping, Egon Loebner, Eo Anne Paulson, Geoffrey K.. Sag, and Thomas Wasow Computer R

PROCESSING ENGLISH WITH A GENERALIZED PHRASE STRUCTURE GRAMMAR Jean Mark Gawron, Jonathan King, John Lamping, Egon Loebner,

Eo Anne Paulson, Geoffrey K Pullum, Ivan A Sag, and Thomas Wasow

Computer Research Center Hewlett Packard Company

1501 Page Mill Road Palo Alto, CA 94304

ABSTRACT

This paper describes a natural language

processing system implemented at Hewlett-Packard's

Computer Research Center The system's main

components are: a Generalized Phrase S t r u c t u r e

Grammar (GPSG); a top-down parser; a logic

t r a n s d u c e r that outputs a f i r s t - o r d e r logical

representation; and a "disambiguator" that uses

sortal information to convert "normal-form"

f i r s t - o r d e r logical expressions into the q u e r y

language f o r HIRE, a relational database hosted in

the SPHERE system We argue that theoretical

developments in GPSG syntax and in Montague

semantics have specific advantages to b r i n g to this

domain of computational linguistics The s y n t a x

and semantics of the system are t o t a l l y

domain-independent, and thus, in p r i n c i p l e ,

h i g h l y portable We discuss the prospects f o r

extending domain-independence to the lexical

semantics as well, and thus to the logical semantic

representations

I INTRODUCTION

This paper is an i n t e r i m progress report on

l i n g u i s t i c research carried out at Hewlett-Packard

Laboratories since the summer of 1981 The

research had three goals: (1) demonstrating the

computational t r a c t a b i l i t y of Generalized Phrase

S t r u c t u r e Grammar (GPSG), (2) implementing a

GPSG system covering a large fragment of English,

and (3) establishing the f e a s i b i l i t y of using GPSG

f o r interactions with an inferencing knowledge

base

Section 2 describes the general a r c h i t e c t u r e

of the system Section 3 discusses the grammar

and the lexicon A b r i e f dicussion of the parsing

technique used in found in Section 4 Section 5

discusses the semantics of the system, and Section

6 presents ~ detailed example of a p a r s e - t r e e

complete with semantics S o m e typical examples

that the system can handle are given in the

Appendix

The system is based on recent developments

in syntax and semantics, reflecting a modular view

in which grammatical s t r u c t u r e an~ abstract logical

s t r u c t u r e have independent status The

understanding of a sentence occurs in a number of

stages, d i s t i n c t from each other and governed by

d i f f e r e n t principles of organization We are

opposed to the idea that language understanding

can be achieved without detailed s y n t a c t i c analysis There is, of course, a massive pragmatic component to human l i n g u i s t i c interaction But we hold that pragmatic inference makes use of a logically p r i o r grammatical and semantic analysis This can be f r u i t f u l l y modeled and exploited even in the complete absence of any modeling of pragmatic inferencing c a p a b i l i t y However, this does not entail an i n c o m p a t i b i l i t y between o u r work and research on modeling

interaction directly= Ultimately, a successful language understanding system wilt require both kinds of research, combining the advantages of precise, g r a m m a r - d r i v e n analysis of utterance

s t r u c t u r e and pragmatic inferencing based on discourse s t r u c t u r e s and knowledge of the world

We stress, however, that our concerns at this stage do not extend beyond the specification of a system that can e f f i c i e n t l y e x t r a c t literal meaning from isolated sentences of a r b i t r a r i l y complex grammatical s t r u c t u r e Future systems will e x p l o i t the literal meaning thus extracted in more ambitious applications that i n v o l v e pragmatic reasoning and discourse manipulation

The system embodies two features t h a t simultaneously promote e x t e n s i b i l i t y , f a c i l i t a t e modification, and increase efficiency The f i r s t is that its grammar is c o n t e x t - f r e e in the informal sense sometimes ( r a t h e r misleadingly) used in discussions of the autonomy of grammar and pragmatics: the syntactic rules and the semantic translation rules are independent of the specific application domain Our rules are not devised ad hoc with a p a r t i c u l a r application or t y p e of interaction in mind Instead, they are motivated

by recent theoretical developments in natural language s y n t a x , and evaluated by the usual linguistic canons of simplicity and g e n e r a l i t y No changes in the knowledge base or other exigencies

d e r i v i n g from a p a r t i c u l a r context of application can introduce a problem f o r the grammar (as

d i s t i n c t , of course, from the lexicon)

The second relevant feature is that the grammar i r the- system is c o n t e x t - f r e e in the sense

of formal language t h e o r y This makes the extensive mathematical l i t e r a t u r e on c o n t e x t - f r e e phrase s t r u c t u r e grammars (CF-PSG's) d i r e c t l y relevant to the e n t e r p r i s e , and permits utilization

of all the well-known techniques f o r the computational implementation of c o n t e x t - f r e e grammars It might seem anachronistic to base a language understanding system on c o n t e x t - f r e e

p a r s i n g As Pratt (1975, 423) observes: " I t is

fashionable these days to want to avoid all

reference to c o n t e x t - f r e e grammars beyond w a r n i n g

s t u d e n t s t h a t t h e y are u n f i t f o r computer

consumption as f a r as computational l i n g u i s t i c s is

c o n c e r n e d " Moreover, w i d e l y accepted arguments

have been given in the l i n g u i s t i c s l i t e r a t u r e to the

effect t h a t some human languages are not even

weakly c o n t e x t - f r e e and t h u s cannot p o s s i b l y be

described by a CF-PSG However, Gazdar and

Pullum (1982) answer all of these arguments,

showing t h a t t h e y are e i t h e r formally i n v a l i d or

e m p i r i c a l l y u n s u p p o r t e d or both It seems

appropriate, t h e r e f o r e , to take a renewed i n t e r e s t

in the p o s s i b i l i t y of CF-PSG d e s c r i p t i o n of human

languages, both in computational l i n g u i s t i c s and in

l i n g u i s t i c research g e n e r a l l y

2 COMPONENTS OF THE SYSTEM

The l i n g u i s t i c basis of the GPSG l i n g u i s t i c

system resides in the w o r k reported in Gazdar

(1981, 1982) and Gazdar, Pullum, and Sag (1981) 1

These papers argue on empirical and theoretical

g r o u n d s t h a t c o n t e x t - f r e e n e s s is a desirable

c o n s t r a i n t on grammars It c l e a r l y would not be

so desirable, however, if (1) it led to lost

generalizations or (2) it resulted in an

unmanageable number of rules in the grammar

Gazdar (1982) proposes a way of simultaneously

a v o i d i n g these two problems L i n g u i s t i c

generalizations can be captured in a c o n t e x t - f r e e

grammar w i t h a metagrammor, i.e a h i g h e r - l e v e l

grammar t h a t generates the actual grammar as its

language The metagrammar has two kinds of

statements:

(1) Rule schemata These are

basically like o r d i n a r y rules, except t h a t

they contain variables r a n g i n g o v e r

categories and features

(2) Metarules These are implicational

statements, w r i t t e n in the form ===>B,

which c a p t u r e relations between rules A

metarule ===>t~ is i n t e r p r e t e d as s a y i n g ,

" f o r e v e r y r u l e t h a t is an i n s t a n t i a t i o n of

the schema =, there is a c o r r e s p o n d i n g rule

of form [5." Here 13 will be @(~), where 8

i s s o m e mapping specified p a r t l y b y the

general t h e o r y of grammar and p a r t l y in

the metarule formulation For instance,

it is taken to be p a r t of the t h e o r y of

grammar t h a t @ preserves unchanged the

subcategorization ( r u l e name) features of

rules (cf below)

The GPSG system also assumes the

R u l e - t o - R u l e Hypothesis, f i r s t advanced by

Richard Montague, which requires t h a t each

s y n t a c t i c rule be associated with a single semantic

I See also Gazdar, Pullum, Sag, and Wasow

(1982) f o r some f u r t h e r discussion and comparison

w i t h other work in the l i n g u i s t i c l i t e r a t u r e

t r a n s l a t i o n rule The s y n t a x - s e m a n t i c s match is realized as follows: each rule is a t r i p l e c o n s i s t i n g

of a r u l e name, a s y n t a c t i c statement (~ormally a local condition on node a d m i s s i b i l i t y ) , and a semantic t r a n s l a t i o n , s p e c i f y i n g how the

h i g h e r - o r d e r logic representations of the d a u g h t e r nodes combine to yield the c o r r e c t t r a n s l a t i o n f o r the mother =

The present GPSG system has f i v e components :

1 Grammar

a Lexicon

b Rules and Metarules

2 Parser and Grammar Compiler

3 Semantics Handler

4 Disambiguator

5 HIRE database

3 GRAMMAR AND LEXICON The grammar t h a t has been implemented t h u s

f a r is o n l y a subset of a much l a r g e r GPSG grammar t h a t we have defined on paper It nevertheless describes a broad sampling of the basic c o n s t r u c t i o n s of English, i n c l u d i n g a v a r i e t y

of prepositional phrase c o n s t r u c t i o n s , n o u n - n o u n compounds, the a u x i l i a r y system, g e n i t i v e s , questions and relative clauses, passives, and

e x i s t e n t i a l sentences

Each e n t r y in the lexicon contains two kinds

of information about a lexical item, s y n t a c t i c and semantic The s y n t a c t i c p a r t of an e n t r y consists

of a s y n t a c t i c f e a t u r e specification; this includes,

inter alia, information about any i r r e g u l a r morphology the item may have, and what is known

in the l i n g u i s t i c l i t e r a t u r e as strict subcategorization information In o u r terms the

l a t t e r is information l i n k i n g lexical items of a

p a r t i c u l a r category to specific environments in which t h a t category is i n t r o d u c e d by phrase

s t r u c t u r e rules Presence in the lexical e n t r y f o r

an item I of the feature R (where R is the name

of a rule) indicates t h a t / may appear in

s t r u c t u r e s admitted by R, and absence indicates

t h a t it may not

The semantic information in a lexical e n t r y is sometimes simple, d i r e c t l y l i n k i n g a lexical item with some HIRE predicate or relation With verbs

or p r e p o s i t i o n s , there is also a specification of what case roles to associate w i t h p a r t i c u l a r arguments (cf below f o r discussion of case roles) Expressions t h a t make a complex logical

c o n t r i b u t i o n to the sentence in which t h e y appear witl in general have complicated t r a n s l a t i o n s

T h u s every has the t r a n s l a t i o n -

2 T h e r e is a theoretical issue here about

w h e t h e r semantic t r a n s l a t i o n rules need to be stipulated f o r each s y n t a c t i c rule or w h e t h e r t h e r e

is a general way of p r e d i c t i n g t h e i r form See Klein and Sag (t981) f o r an attempt to develop the

l a t t e r view, which is not at p r e s e n t implemented

in o u r system

(LAMBDA P (LAMBDA Q ((FORALL X (P X ) )

> (Q x ) ) ) ) , This indicates t h a t it denotes a f u n c t i o n which

takes as argument a set P, and r e t u r n s the set of

p r o p e r t i e s t h a t are t r u e of all members of t h a t set

(cf below f o r s l i g h t l y more detailed d i s c u s s i o n )

A t y p i c a l rule looks like t h i s :

<VPI09: V] -> V N]! N!I2: ((V N!!2) N!!)>

The exclamation marks here are o u r notation f o r

the bars in an X - b a r category system (See

J a c k e n d o f f (1977) f o r a t h e o r y of this

t y p e - - t h o u g h one which d i f f e r s on points of detail

from o u r s ) The rule has the form <a: b: c>

Here a is the name 'VP109'; b is a condition t h a t

will admit a node labeled ' V ! ' if it has t h r e e

d a u g h t e r nodes labeled r e s p e c t i v e l y 'V' ( v e r b ) ,

' N i t ' (noun phrase at the second bar l e v e l ) , and

' N I ! ' (the numeral 2 being merely an index to

permit reference to a specific symbol in the

semantics, t h e metarules, and t h e rule compiler,

and is not a p a r t of the category l a b e l ) ; and c is

a semantic t r a n s l a t i o n rule s t a t i n g t h a t the V

c o n s t i t u e n t translates as a f u n c t i o n expression

t a k i n g as its argument the t r a n s l a t i o n of the

second N ! ! , the result being a f u n c t i o n expression

to be applied to the t r a n s l a t i o n of the f i r s t N ! !

By a general convention in the theory of

grammar, the rule name is one of the feature

values marked on the lexical head of any rule that

introduces a lexical category (as this one

introduces V) Only verbs marked with that

feature value satisfy this rule For example, if we

include in the lexicon the word give and assign to

it the feature VPI09, then this rule would generate

the verb phrase gave Anne a job

A t y p i c a l metarule is the passive metarule,

which looks like this ( i g n o r i n g semantics):

< P A S : <V! -> V NI! W > => <V! -> V [ P A S ] W > >

W is a s t r i n g v a r i a b l e r a n g i n g o v e r zero or more

c a t e g o r y symbols The metarule has the form <N:

<A> => <B>>, where N is a name and <A> and <B >

are schemata t h a t have rules as t h e i r instantiations

when a p p r o p r i a t e s u b s t i t u t i o n s are made f o r the

free v a r i a b l e s This metarule says t h a t f o r e v e r y

rule t h a t expands a v e r b phrase as v e r b followed

by noun phrase followed by a n y t h i n g else

( i n c l u d i n g n o t h i n g else), there is a n o t h e r rule that

expands v e r b phrase as v e r b with passive

morphology followed by w h a t e v e r followed the noun

phrase in the given rule The metarule PAS would

apply to grammar rule VP109 given above, y i e l d i n g

the rule:

<VP109: V! -> V[PAS] N { ! >

As we noted above, the rule number f e a t u r e is

p r e s e r v e d here, so we get Anne was given a job,

where the passive v e r b phrase is given a job,

but not *Anne was hired a job 3

Passive sentences are thus analyzed d i r e c t l y , and not reduced to the form of active sentences in

t h e course of being analyzed, in the way t h a t is

f a m i l i a r from w o r k on t r a n s f o r m a t i o n a l grammars and on ATN's However, this d o e s not mean t h a t

no relation between passives and t h e i r active

c o u n t e r p a r t s is expressed in the system, because

t h e r u l e s f o r a n a l y z i n g passives are in a sense

d e r i v a t i v e l y defined on the basis o f ' rules f o r

a n a l y z i n g actives

More d i f f i c u l t than t r e a t i n g passives and the like, and often cited as l i t e r a l l y impossible w i t h i n a

c o n t e x t - f r e e grammar'," is t r e a t i n g constructions like questions and r e l a t i v e clauses The a p p a r e n t

d i f f i c u l t y resides in the fact t h a t in a question like

Which employee has Personnel reported that Anne thinks has performed outstandingly?, the p o r t i o n

b e g i n n i n g with the t h i r d word must c o n s t i t u t e a

s t r i n g analyzable as a sentence e x c e p t t h a t at some

p o i n t it must lack a t h i r d person s i n g u l a r noun phrase in a position where such a noun phrase could otherwise have o c c u r r e d If it lacks no noun phrase, we get ungrammatical s t r i n g s of the

t y p e *Which employee has Personnel reported that Anne thinks Montague has performed outstandingly? If it lacks a noun phrase at a position where the v e r b agreement indicates something o t h e r than a s i n g u l a r one is r e q u i r e d ,

we get ungrammaticalities like *Which employee has Personnel reported that Anne thinks have performed outstandingly? The problem is thus one of g u a r a n t e e i n g a grammatical dependency across a c o n t e x t t h a t may be a r b i t r a r i l y wide, while keeping the grammar c o n t e x t - f r e e The

t e c h n i q u e used is i n t r o d u c e d into the l i n g u i s t i c

l i t e r a t u r e by Gazdar (1981) It involves an augmentation of the nonterminal v o c a b u l a r y of the grammar t h a t permits c o n s t i t u e n t s with "gaps" to

be t r e a t e d as not b e l o n g i n g to the same category

as similar constituents w i t h o u t gaps This would

be an unwelcome and inelegant enlargement of the grammar if it had to be done by means of case-by-case s t i p u l a t i o n , but again the use of a metagrammar avoids this Gazdar (1981) proposes

a new set of syntactic categories of the form a/B,

where ~ and 15 are categories from the basic nonterminal v o c a b u l a r y of the grammar These are called slash categories A slash category e/B may

be t h o u g h t of as r e p r e s e n t i n g a c o n s t i t u e n t of

c a t e g o r y = with a missing i n t e r n a l occurrence of !5

We employ a method of i n t r o d u c i n g slash categories

t h a t was suggested by Sag (1982): a metarule

s t a t i n g t h a t f o r e v e r y rule i n t r o d u c i n g some B

u n d e r = t h e r e is a parallel rule i n t r o d u c i n g 15/~

u n d e r =/~ In o t h e r words, any c o n s t i t u e n t can have a gap of t y p e ~" if one of its d a u g h t e r

c o n s t i t u e n t s does too Wherever this would lead to

a d a u g h t e r c o n s t i t u e n t with the label [/~' in some

3 ~ regard was given a job not as a passive

v e r b phrase itself b u t as a v e r b phrase c o n t a i n i n g the v e r b be plus a passive v e r b phrase containing

given and a job

4 See Pullum and Gazdar (1982) for references

r u l e , a n o t h e r metarule allows a parallel rule

w i t h o u t the ~'/;r, and t h e r e f o r e defines rules t h a t

allow f o r actual g a p s - - i e , missing c o n s t i t u e n t s

In t h i s way, complete sets of rules f o r d e s c r i b i n g

the u n b o u n d e d dependencies f o u n d in i n t e r r o g a t i v e

and r e l a t i v e clauses can r e a d i l y be w r i t t e n Even

l o n g - d i s t a n c e agreement facts can be (and are)

c a p t u r e d , since the m o r p h o s y n t a c t i c features

r e l e v a n t to a specific case of agreement are

p r e s e n t in the feature composition of any given ~'

4 PARSING

The system is i n i t i a l i z e d b y e x p a n d i n g out

the g r a m m a r T h a t is, t i l e metarules are applied

to the rules to produce the f u l l rule set, which is

then compiled and used b y the parser Metarules

are not consulted d u r i n g the process of p a r s i n g

One might well wonder about the possible benefits

of the o t h e r a l t e r n a t i v e : a p a r s e r t h a t made the

m e t a r u l e - d e r i v e d rules to o r d e r each time t h e y

were needed, instead of c o n s u l t i n g a precompiled

l i s t T h i s p o s s i b i l i t y has been explored by Kay

(1982) Kay draws an analogy between metarules

and phonological rules, modeling both by means of

f i n i t e state t r a n s d u c e r s We believe t h a t t h i s line

is w o r t h p u r s u i n g ; however, the GPSG system

c u r r e n t l y operates off a precompiled set of rules

A p p l i c a t i o n of ten metarules to f o r t y basic

rules yielded 283 grammar rules in the 1/1/82

version of the GPSG system Since then the

grammar has been expanded somewhat, t h o u g h the

c u r r e n t version is still u n d e r g o i n g some

d e b u g g i n g , and the number of rules is unstable

T h e size of the g r a m m a r - p l u s - m e t a r u l e s system

grows b y a f a c t o r of f i v e or six t h r o u g h the rule

compilation The great practical advantage of

using a m e t a r u l e - i n d u c e d grammar is, t h e r e f o r e ,

t h a t the w o r k of designing and r e v i s i n g the system

of l i n g u i s t i c rules can proceed on a body of

statements t h a t is u n d e r t w e n t y p e r c e n t of the size

it would be if it were formulated as a simple list of

c o n t e x t - f r e e rules

The system uses a standard t y p e of

t o p - d o w n p a r s e r w i t h no Iookahead, augmented

s l i g h t l y to p r e v e n t it from looking f o r a given

c o n s t i t u e n t s t a r t i n g in a given spot more than

once It produces, in parallel, all legal parse

trees f o r a sentence, w i t h semantic t r a n s l a t i o n s

associated w i t h each node

5 SEMANTICS

The semantics h a n d l e r uses the t r a n s l a t i o n

rule associated w i t h a node to c o n s t r u c t its

semantics from the semantics of its d a u g h t e r s

T h i s c o n s t r u c t i o n makes crucial use of a procedure

t h a t we call Cooper storage ( a f t e r Robin Cooper;

see below) In the s p i r i t of c u r r e n t research in

formal semantics, each s y n t a c t i c c o n s t i t u e n t is

associated d i r e c t l y with a single logic expression

(modulo Cooper Storage), r a t h e r than any program

or p r o c e d u r e for p r o d u c i n g such an expression

O u r semantic analysis thus embraces the p r i n c i p l e

of " s u r f a c e c o m p o s i t i o n a l i t y " The semantic

representations d e r i v e d at each node are r e f e r r e d

to as the Logical Representation ( L R ) The disambiguator p r o v i d e s the c r u c i a l

t r a n s i t i o n from LR to H I R o E queries; the disambiguator uses information about the sort, or

domoin of definition, of v a r i o u s terms in the logical

r e p r e s e n t a t i o n One of the most i m p o r t a n t

f u n c t i o n s of the disambiguator is to eliminate parses t h a t do not make sense in the conceptual scheme of HIRE

HIRE is a relational database w i t h a certain amount of i n f e r e n c i n 9 c a p a b i l i t y It is implemented

in SPHERE, a database system which is a descendant of FOL ( d e s c r i b e d in Weyhrauch (1980)) Many of the relation-names o u t p u t by the d i s a m b i g u a t o r are d e r i v e d relations defined b y axioms in SPHERE The SPHERE e n v i r o n m e n t was important f o r t h i s a p p l i c a t i o n , since it was essential to have something t h a t could process

f i r s t - o r d e r logical o u t p u t , and SPHERE does j u s t

t h a t A noticeable recent t r e n d in database t h e o r y has been a move t o w a r d an i n t e r d i s c i p l i n a r y comingling of mathematical logic and relational database t e c h n o l o g y (see especially Gallaire and

M i n k e r (1978) and Gallaire, M i n k e r and Nicolas ( 1 9 8 ] ) ) We regard it as an important fact about the GPSG system t h a t links computational

l i n g u i s t i c s to f i r s t - o r d e r logical representation

j u s t as the w o r k r e f e r r e d to above has linked

f i r s t - o r d e r logic to relational database t h e o r y We believe t h a t SPHERE offers promising prospects f o r

a knowledge representation system t h a t is

p r i n c i p l e d and general in the way t h a t we have

t r i e d to e x e m p l i f y in o u r s y n t a c t i c and semantic rule system Filman, Lamping and Montalvo (]982)

p r e s e n t details of some capabilities of SPHERE t h a t

we have not as yet e x p l o i t e d in o u r w o r k ,

i n v o l v i n g the use of multiple c o n t e x t s to r e p r e s e n t

v i e w p o i n t s , beliefs, and modalities, which are

g e n e r a l l y regarded as i n s u p e r a b l e s t u m b l i n g - b l o c k s

to f i r s t - o r d e r logic approaches

T h u s f a r the l i n g u i s t i c w o r k we have described has been in keeping w i t h GPSG presented in the papers cited above However two semantic innovations have been i n t r o d u c e d to

f a c i l i t a t e the disambiguator's t r a n s l a t i o n from LR to

a HIRE q u e r y As a r e s u l t the l i n g u i s t i c system version of LR has two new p r o p e r t i e s :

(1) The intensional logic of the p u b l i s h e d work was set aside and LR was designed to be an extensional f i r s t - o r d e r language A l t h o u g h

c o n s t i t u e n t t r a n s l a t i o n s b u i l t up on the way to a root node may be s e c o n d - o r d e r , the system- maintains f i r s t - o r d e r r e d u c i b i l i t y T h i s

r e d u c i b i l i t y is i l l u s t r a t e d by the f o l l o w i n g analysis

of noun phrases as s e c o n d - o r d e r p r o p e r t i e s (essentially the analysis of Montague ( ] 9 7 0 ) ) For example, the p r o p e r name Egon and the q u a n t i f i e d noun phrase every opplicant are both t r a n s l a t e d as sets of p r o p e r t i e s :

77

Egon = LAMBDA P (P EGON)

E v e r y a p p l i c a n t = LAMBDA P (FORALL X

( ( A P P L I C A N T X) > (P X ) ) )

Egon is t r a n s l a t e d as the set of p r o p e r t i e s

t r u e of Egon, and every applicant, as the set of

p r o p e r t i e s t r u e of all applicants Since basic

predicates in the logic are f i r s t - o r d e r , n e i t h e r of

• a r g u m e n t of any basic predicate; instead the

argument is some u n i q u e e n t i t y - l e v e l v a r i a b l e

which is later bound to the q u a n t i f i e r - e x p r e s s i o n

b y q u a n t i f y i n g in T h i s t e c h n i q u e is essentially

One advantage of t h i s method of " d e f e r r i n g " the

i n t r o d u c t i o n into the i n t e r p r e t a t i o n process of

phrases w i t h q u a n t i f i e r meanings is that it allows

f o r a n a t u r a l , n o n s y n t a c t i c treatment of scope

ambiguities A n o t h e r is t h a t w i t h a logic limited to

f i r s t - o r d e r predicates, t h e r e is still a natural

t r e a t m e n t f o r coordinated noun phrases of

a p p a r e n t l y heterogeneous semantics, such as Egon

and every applicant

(2) HIRE represents events as objects All

objects in the knowledge base, i n c l u d i n g events,

belong to various sorts For o u r purposes, a sort

is a set HIRE relations are declared as p r o p e r t i e s

of entities w i t h i n p a r t i c u l a r sorts For example,

t h e r e is an employment sort, c o n s i s t i n g of various

p a r t i c u l a r employment events, and an

employment.employee relation as well as

employment organization and employment.manager

relations More conventional relations, like

employee.manager are defined as joins of the basic

some f a i r l y obvious connections between v e r b s and

events (between, say, the v e r b work and events

of employment), and to represent d i f f e r e n t

relations between a v e r b and its arguments as

d i f f e r e n t f i r s t - o r d e r relations between an event

and its p a r t i c i p a n t s A l t h o u g h the lexical

treatment sketched here is c l e a r l y domain

d e p e n d e n t (the English v e r b work doesn't

necessarily i n v o l v e employment e v e n t s ) , it was

chosen p r i m a r i l y to s i m p l i f y the ontology of a f i r s t

implementation As an a l t e r n a t i v e , one might

consider associating work w i t h events of a sort

labor, one of whose subsorts was an employment

e v e n t , d e f i n i n g employments as those labors

associated w i t h an o r g a n i z a t i o n

Whichever choice one makes about the basic

e v e n t - t y p e s of v e r b s , the mapping from verbs to

HIRE relations cannot be d i r e c t Consider a

sentence like Anne work5 for Egon The HIRE

employment.manager relation of a p a r t i c u l a r

employment event and a p a r t i c u l a r manager, and

t h e employment.employee relation of t h a t same

e v e n t and ,~knl,~ Yet where Egon in t h i s example

is picked out w i t h the employment manager

relation, the sentence Anne worl<s for HP will need

to pick out HP w i t h the employment.organization

relation I n o r d e r to accomodate t h i s

m a n y - t o - m a n y mapping between a v e r b and

p a r t i c u l a r relations in a knowledge base, the

lexicon stipulates special relations t h a t l i n k a

v e r b to its eventual arguments Following Fillmore

(1968), these mediating relations are called case roles

The disambiguator narrows the case roles down to specific knowledge base relations To take a simple example, Anne works for HP has a logical representation r e d u c i b l e to:

(EXISTS SIGMA (AND (EMPLOYMENT SIGMA)

(AG SIGMA ANNE) (LOC SIGMA H P ) ) ) Here SIGMA is a v a r i a b l e o v e r s i t u a t i o n s or event

i n s t a n t i a t i o n s , s The formula may be read, " T h e r e

is an employment-situation whose A g e n t is A n n e and whose Location is H P " The lexical e n t r y f o r

work supplies the information t h a t its subject is an

A g e n t and its complement a Location T h e disambiguator now needs to f u r t h e r specify the case roles as HIRE relations It does t h i s by

t r e a t i n g each atomic formula in the expression locally, using the f a c t t h a t A n n e is a person in

o r d e r to i n t e r p r e t AG, and the fact t h a t HP is

an o r g a n i z a t i o n in o r d e r to i n t e r p r e t LOC In t h i s case, it i n t e r p r e t s the AG role as

employment.employee and the LOC role as

employment.organization

T h e advantages of u s i n g the roles in Logical Representation, r a t h e r than going d i r e c t l y to predicates in a knowledge base, include (1) the

a b i l i t y to i n t e r p r e t at least some prepositional phrases, those known as a d j u n c t s , w i t h o u t

s u b c a t e g o r i z i n g v e r b s specially f o r them, since the case role may be supplied e i t h e r b y a v e r b or a

p r e p o s i t i o n (2) the option of i n t e r p r e t i n g 'vague' v e r b s such as have and give using case

roles w i t h o u t e v e n t t y p e s These v e r b s , t h e n , become " p u r e l y " relational For example, the representation of Egon gave Montague a job would be:

(EXISTS SIGMA (AND ((SO EGON) SIGMA)

((POS MONTAGUE) SIGMA) (EMPLOYMENT S I G M A ) ) ) Here SO 'source' w i l l pick out the same

employment.manager relation it did in the example above; and POS 'possession' is the same relation as

t h a t associated w i t h have Here the s i t u a t i o n - t y p e

is supplied by the t r a n s l a t i o n of the noun job It

is important to realize t h a t t h i s representation is

d e r i v e d w i t h o u t g i v i n g the noun phrase a job any special treatment The lexical e n t r y f o r give

contains the information t h a t the subject is the source of the d i r e c t object, and the d i r e c t object the possession of the i n d i r e c t object If t h e r e were lamps in o u r knowledge base, the d e r i v e d representation of Egon gave Montague a lamp would simply be the above formula w i t h the predicate

lamp replacing employment The possession relation would hold between Montague and some

5 O u r w o r k in t h i s domain has been i n f l u e n c e d

b y the recent papers of Barwise and Perry on

" s i t u a t i o n semantics"; see e c Barwise and P e r r y (1982))

lamp, and the disambiguator would r e t r i e v e

w h a t e v e r knowledge-base relation kept t r a c k of

such matters

Two active research goals o f the c u r r e n t

project are to give all lexical entries domain

i n d e p e n d e n t r e p r e s e n t a t i o n s , and to make all

knowledge base-specific predicates and relations

the e x c l u s i v e p r o v i n c e of the disambiguator One

i m p o r t a n t means to t h a t end is case roles, which

allow us a level of a b s t r a c t , p u r e l y " l i n g u i s t i c "

relations to mediate between logical representations

and HIRE queries A n o t h e r is the use of general

e v e n t types such as labor, to replace e v e n t - t y p e s

specific to HIRE, such as employments The case

roles maintain a separation between the domain

representation language and LR Insofar as t h a t

separation is achieved, then absolute p o r t a b i l i t y

of the system, up to and i n c l u d i n g the l e x i c o n , is

an attainable goal

Absolute p o r t a b i l i t y o b v i o u s l y has immediate

practical benefits f o r any system t h a t expects to

handle a large fragment of English, since the

e f f o r t in moving from one application to a n o t h e r

will be limited to " t u n i n g " the disambiguator to a

new o n t o l o g y , and adding "specialized" v o c a b u l a r y

The actual rules g o v e r n i n g the p r o d u c t i o n of

f i r s t - o r d e r logical representations make no

reference to the facts of HIRE The question

remains of j u s t how portable the c u r r e n t lexicon

is; the answer is t h a t much of it is already domain

i n d e p e n d e n t Q u a n t i f i e r s like e v e r y (as we saw in

the discussion of NP semantics) are expressed as

logical constants; v e r b s like give are expressed

e n t i r e l y in terms of the case relations t h a t hold

among t h e i r arguments Verbs like w o r k can be

abstracted away from the domain by a simple

e x t e n s i o n The obvious goal is to t r y to g i v e

domain independent representations to a core

v o c a b u l a r y of English t h a t could be used in a

v a r i e t y of application domains

6 AN EXAMPLE

We shall now g i v e a s l i g h t l y more detailed

i l l u s t r a t i o n of how the s y n t a x and compositional

semantics rules w o r k We are still s i m p l i f y i n g

c o n s i d e r a b l y , since we have selected an example

where rote frames are not i n v o l v e d , and we are

not employing features on nodes Here we have

the grammar of a t r i v i a l subset of English:

<$1: S -> NP VP: (NP Vp)>"

<NPI: NP -> DET N: (DET N)>

<VPI: VP -> V NP: i V NP)>

<VP2: VP -> V A: A>

Suppose t h a t the lexicon associated w i t h the above

rules is:

< e v e r y : D E T : (LAMBDA P (LAMBDA Q

(FORALL X ((P X) IMPLIES (Q X ) ) ) ) ) >

<applicant: N: APPLICANT>

< i n t e r v i e w e d : V [ ( R U L E V P 1 ) ] : INTERVIEW>

<Bill: NP: (LAMBDA P (P B I L L ) ) >

<is: V [ ( R U L E MP2)]: (BE)>

<competent: A: (LAMBDA Y

( E X P E R T L E V E L HIGH Y ) ) >

The s y n t a x of a lexical e n t r y is <L: C: T>, where

L is t h e spelling of the item, C is its grammatical category and f e a t u r e specification ( i f o t h e r than

the d e f a u l t set) and T is its translation into LR

Consider how we assign an LR to a sentence

like E v e r y applicant is competent The t r a n s l a t i o n

of e v e r y supplies most of the s t r u c t u r e of the

u n i v e r s a l q u a n t i f i c a t i o n needed in LR I t represents a f u n c t i o n from p r o p e r t i e s to f u n c t i o n s from p r o p e r t i e s to t r u t h values, so when applied

to applicant it y i e l d s a c o n s t i t u e n t , namely e v e r y applicant, which has one of the p r o p e r t y slots

f i l l e d , and represents a f u n c t i o n from p r o p e r t i e s

to t r u t h - v a l u e s ; it is:

(LAMBDA P (FORALL X ( ( A P P L I C A N T X) IMPLIES (P X ) ) ) )

T h i s f u n c t i o n can now be applied to the f u n c t i o n

denoted b y competent, i.e

( LAMBDA Y ( E X P E R T L E V E L HIGH Y ) )

T h i s y i e l d s : (FORALL X ( ( A P P L I C A N T X ) IMPLIES

(LAMBDA Y ( E X P E R T L E V E L HIGH Y ) ) X ) ) And a f t e r one more lambda-conversion, we have:

( FORALL X ( ( A P P L I C A N T X ) IMPLIES

( E X P E R T L E V E L HIGH X ) ) ) Fig 1 shows one parse tree t h a t would be generated b y the above rules, t o g e t h e r w i t h its logical t r a n s l a t i o n The sentence is B i l l

i n t e r v i e w e d e v e r y applicant The complicated

t r a n s l a t i o n of the VP is necessary because INTERVIEW is a one-place predicate t h a t takes an

e n t i t y - t y p e argument, not the t y p e of f u n c t i o n

t h a t e v e r y applicant denotes We t h u s defer combining the NP t r a n s l a t i o n w i t h the v e r b b y using Cooper storage A t r a n s l a t i o n w i t h a stored

NP is represented above in a n g l e - b r a c k e t s Notice

t h a t at the S node the NP e v e r y applicant is still

stored, b u t the subject is not stored It has

d i r e c t l y combined w i t h the VP, by t a k i n g the VP

as an argument INTERVIEW is itself a two-place predicate, b u t one of its argument places has been

f i l l e d by a p l a c e - h o l d i n g variable, X1 T h e r e is th~Js ~ o n l y one slot left The t r a n s l a t i o n can now

be completed via the operations of Storage Retrieval and lambda c o n v e r s i o n F i r s t , we s i m p l i f y the p a r t of the semantics t h a t i s n ' t in storage:

79

Fig 1 A typical parse tree

S

<((LAMBDA P (P BILL))(INTERVIEW X1)),

<(LAMBDA P (FORALL X ((APPLICANT X) IMPLIES (P X ) ) ) ) >>

NP ((LAMBDA P (P B I L L ) ) )

VP

<(INTERVIEW X1) (LAMBDA P (FORALL X ((APPLICANT X) IMPLIES (P X ) ) ) ) >

INTERVIEW

I

interviewed

NP (LAMBDA P (FORALL X ((APPLICANT X) IMPLIES (P X ) ) ) )

LAMBDA Q (LAMBDA P (FORALL X ((Q X) IMPLIES (P X ) ) ) ) every

((LAMBDA P (P BILL))(INTERVIEW X1)) :>

((INTERVIEW X l ) BILL)

The function (LAMBDA P (P BILL)) has been

evaluated with P set to the value (INTERVIEW

X1); this is a conventional lambda-conversion

The rule for storage retrieval is to make a

one-place predicate of the sentence translation by

lambda-binding the placeholding variable, and then

to apply the NP translation as a function to the

result The S-node translation above becomes:

((LAMBDA P

(FORALL X

((APPLICANT X) IMPLIES (P X ) ) ) )

(LAMBDA X1 ((INTERVIEW X1) B I L L ) ) )

[lambda-conversion] = = >

(FORALL X ((APPLICANT X) IMPLIES

((LAMBDA X1

((INTERVIEW X1) BILL)) X ) ) )

[lambda-conversion] : : >

(FORALL X ((APPLICANT X) IMPLIES

(((INTERVIEW X) B I L L ) ) ) )

This is the desired final result

7 CONCLUSION What we have outlined is a natural language system that is a direct implementation of a linguistic theory We have argued that in this case the linguistic theory has the special appeal of computational t r a c t a b i l i t y (promoted by its context-freeness), and that the system as a whole offers the hope of a happy marriage of linguistic

computer applications The system's theoretical underpinnings give it compatibility with c u r r e n t research in Generalized Phrase Structure Grammar, and its augmented f i r s t order logic gives it compatibility with a whole body of ongoing research in the field of model-theoretic semantics The work done thus far is only the f i r s t step on the road to a robust and practical natural language processor, but the guiding principle throughout has been e x t e n s i b i l i t y , both of the grammar, and of the applicability to various spheres of computation

ACKNOWLEDGEMENT Grateful acknowledgement is given to two brave souls, Steve Gadol and Bob Kanefsky, who helped give this system some of its c r e d i b i l i t y by implementing the actual hook-up with HIRE Thanks are also due Robert Filman and Bert Raphael for helpful comments on an early version

of this paper And a special thanks is due Richard Weyhrauch, for encouragement, wise advice, and comfort in times of debugging

APPENDIX

This appendix lists some sentences that are

actually translated into HIRE and answered by the

current system Declarative sentences presented

to the system are evaluated with respect with

their t r u t h value in the usual way, and thus also

function as queries

SIMPLE SENTENCES

1 HP employs Egon

2 Egon works for HP

3 HP offered Montague the position

4 HP gave Montague a job

5 Montague got a job from HP

6 Montague's job is at HP

7 HP's offer was to Capulet

8 Montague had a meeting with Capulet

9 Capulet has an offer from Xerox

10 Capulet is competent

IMPERATIVES AND QUESTIONS

11 Find the programmers in CRC

who attended the meeting

12 How many applicants for the

position are there?

13 Which manager interviewed Capulet?

14 Whose job did Capulet accept?

15 Who is a department manager?

16 Is there a LISP programmer

who Xerox hired?

17 Whose job does Montague have?

18 How many applicants

did Capulet interview?

RELATIVE CLAUSES

19 The professor whose student Xerox

hired visited HP

20 The manager Montague met with hired

the student who attended Berkeley

NOUN-NOUN COMPOUNDS

21 Some Xerox programmers visited HP

22 Montague interviewed a job applicant

23 Who are the department managers?

24 How many applicants have a LISP

programming background?

COORDINATION

25 Who did Montague interview and visit?

26 Which department's position did

every programmer and a manager

from Xerox apply for?

PASSIVE AND EXISTENTIAL SENTENCES

27 Egon was interviewed by Montague

28 There is a programmer

who knows LISP in CRC

INFINITIVAL COMPLEMENTS

29 Montague managed to get a job at HP

30 HP failed to hire a programmer

with Lisp programming background

REFERENCES

"Situations and attitudes." Journal of Philosophy 78, 668-692

Cooper, Robin 1 9 7 5 Montague's Semantic Theory and Transformational Syntax

Doctoral dissertation, University of Massachusetts, Amherst

Fillmore, Charles 1968 "The Case for Case."

In Bach, Emmon and Robert Harms

Universals in Linguistic Theory New York: Holt, Rinehart and Winston

Filman, Robert E., John Lamping, and Fanya Nlontalvo 1 9 8 2 "Metalanguage and

presentation at the AAAI National Conference on Artificial Intelligence, Carnegie-Mellon University, Pittsburgh, Pennsylvania

Gallaire, Herv$, and Jack Minker, eds 1978

Logic and Data Bases New York: Plenum Press

Gallaire, Herv$, Jack Minker, and Jean Marie Nicolas, eds 1 9 8 1 Advances in Date Base Theory New York: Plenum Press Gazdar, Gerald 1 9 8 1 "Unbounded Dependencies and Coordinate S t r u c t u r e "

Linguistic Inquiry 12, 155-184

Gazdar, Gerald 1 9 8 2 "Phrase Structure Grammar." In Pauline Jacobson and Geoffrey K Pullum, eds The Nature of Syntactic Representation Dordrecht: D Reidel

Gazdar, Gerald, Geoffrey K Pullum, and Ivan

A Sag In press "Auxiliaries and Related Phenomena." Language

Gazdar, Gerald, Geoffrey K Pullum, Ivan A

"Coordination and Transformational Grammar" Linguistic Inquiry 13

Jackendoff, Ray 1977 ~" Syntax Cambridge: MIT Press

Kay, Martin 1982 "When Metarules are not Metarules." Ms Xerox Palo Alto Research Center

Montague, Richard 1 9 7 0 "The Proper Treatment of Quantification in English."

in Richmond Thomason, ed 1974 Formal Philosophy New Haven: Yale U n i v e r s i t y Press

Pratt, Vaughan R 1 9 7 5 "LINGOL a progress r e p o r t " Advance Papers of the Fourth /nternational Joint Conference on Artificia/ /nte//igence, Tbilisi, Georgia, USSR, 3-8 September 1975 Cambridge, MA: Artificial Intelligence Laboratory 422-428

Pullum, Geoffrey K and Gerald Gazdar

1982 Natural languages and context-free languages Linguistics and phitos.ophy 4 Sag, Ivan A 1982 "Coordination, Extraction,

Grammar." Linguistic Inquiry 13

Weyhrauch, Richard W 1980 "Prolegomena to

a theory of mechanized formal reasoning." Artificial Intelligence, 1, pp 133-170