037_2002_v1_josef ruppenhofer, collin f. baker & charles j. fillmore_collocational information in

8/18/2019 037_2002_V1_Josef Ruppenhofer, Collin F. Baker & Charles J. Fillmore_Collocational Information In

1/11

REFORTS ON

LEXICOGRAPHICAL

AN D

LEXICOLOGICAL

PROJECTS

Collocational Information

in

the

FrameNet Database

JosefRuppenhofer,

Collin

F.

Baker,

and Charles

J.

Fillmore

International

Computer Science Institute

1947CenterSt .

Berkeley, CA

94704-1198,

USA

ftosef,

collinb,

fillmore}@icsi.berkeley.edu

website:

http://framenet.icsi.berkeley.edu/~framenet

Abstract

Th e FrameNet

exical

atabase

yields

nformation

bout collocations

nd

multiword

xpressions n

various

ways.

n

om e

cases

phrasal

units

have

been entered

from th e start as

lexical

entries (write

down). In

other

cases headword+preposition pairs

can

be recognized

as

special

collocations

where

th e preposition

in

question

is

a ecessary

nd

exically

pecified

marker

of

an rgument

of th e

eadword

fond

of, ostile

o) . Nominal

compounds

re

nnotated with espect

o

oun r

pertinative) djective modifiers, om e of which re

analyzable

but

also entrenched

wheel chair, fiscal year). Nouns that

name

ggregates, portions,

ypes,

tc.,

sometimes

hold

lexically specified relations to

their dependents •

•

•

ofgeese). nd event

nouns

frequently

select

th e upport

verbs

which

ermit

them o

nter

nto

redications

file

n

bjection,

nter a

plea).

A

subproject aims

at extracting, as

structured

clusters of

lexical

tems, th e minimal

emantically

central kernel

dependency

graphs

ro m th e set

of

annotations.

Such research

will yield

not

only

ommonplace groupings

(eat:

og, bone)

but

will

ls o

yield

itherto unnoticed

ollocations

within

uch graphs

(answer:

ou,

oor)

where

certain dependency links

within

them are

idiomatic

or

otherwise

lexically

special,

here answer

door.

Cpllocational

information

can

also be

retrieved

by various types ofqueries within our MySQL search tool.

Introduction

Th e FrameNet

research

project

[Baker

et

al.

998;

Fillmore

&

Baker

2001]

s building an

online

lexical

resource that

aims

to

provide,

fo r

a

significant

portion

of

th e

vocabulary

of

contemporary

English,

a

body

of

semantically and

syntactically

annotated

sentences

from

which

reliable information

can be

reported

on th e

valences

or

combinatorial

possibilities of

each

item

included.

The project

uses

a descriptive

model

based on semantic

frames

[Fillmore

1977,

982,

985;

Fillmore

&

Atkins

988]

nd documents

ts

observations

by

means

of

carefully

annotated

attestations

taken

from corpora, each

sentence

annotated

in

respect to a

single

arget

word

with

th e

phrases that

are

n grammatical

onstruction

with t

abelled

according

to their

grammatical

relation to th e target, th e

semantic

role they serve

within

th e

frame

to

which th e target word

belongs,

and

it s

syntactic

phrase type.

Th e FN

database

can serve

both

human and

machine users"

nd

can

function

both as

a

dictionary

and as

a

thesaurus.

As

a

dictionary,

each

lexical

unit

( lemma in a

given

sense)

is

provided

with (1) th e name of

th e

rame

t belongs to and access to a description

of

th e

frame, (2 )

a

definition

(either

original

or

from th e Concise Oxford Dictionary, courtesy of

Oxford

niversity

ress),

3)

alence

escription

hich

ummarizes

he

ttested

combinatorial ossibilities n espect

o oth emantic oles nd he yntactic orm nd

function of th e hrases

that

nstantiate those roles, nd (4) ccess o

annotated

xamples

illustrating

each

syntactic

pattern

found

in

th e

corpus

and th e

kinds

ofsemantic information

they contain. The

semantic

role

annotation

s

done

manually

y

persons

trained n

rame

359


2/11

EURALEX

2 2

PROCEEDINGS

semantic

heory;

he yntactic

nformation

s

dded

utomatically,

nd he

ull alence

descriptions

are

produced

automatically.

It

is

possible

to

consider

th e

database

as

a

thesaurus

by

noting

that lemmas

are

linked

to

th e

semantic frames in

which they

participate,

and frames, in

turn,

are

l inked

both

to

th e full

set

ofwords which instantiate them and to related frames. Frame-to-frame relations include (1)

composition,

y

which

omplex

rame s

ecomposable

nto

ubframes,

ften

n

structured

procedural sequence

(thus,

th e Arraignment frame

is

treated as

a subframe

of

th e

Criminal Justice rame), nd (2 ) nheritance,

by which

a

single rame can be seen

as

n

elaboration of

one

or

more

other frames, with bindings between

th e

inherited semantic

roles

(a s

when

criticize

(in

th e Judgement_Communication frame) can

be

seen

as

inheriting

from

both

th e

Judgement

and

Communication

frames,

requiring

a

binding

between

th e

Speaker

of

th e

Communication

frame

and

th e

Judge

ofthe

Judgement.

The

Fram eN e t

Database

The FN ata are

tored

n

MySQL

database

Fillmore t

al . 001] which s

asically

divided nto

w o

alves,

ne

epresenting

he rames, he

rame

lements,

he emmas

connected

with them and th e

relations

among

them (shown

in

Fig. ),

nd th e

other

(not

shown)

representing

th e

corpus

sentences

and

th e

labels

attached to them,

marking

phrases

as

instantiations of given frame elements, phrase types and grammatical

functions, etc. This

division

corresponds

to

th e

tw o

main

software

tools

used

in

FN

work, th e

Frame

Editor

and

th e

Annotation

tool,

both

ofwhich

will

be

demonstrated

in

th e

FrameNet

demonstration.

Lemma

CreatedD4e

Crct f edBy

Q P a tO S p ch _ R e l

&ocked

STInheril

QP* e n t ST _Re l

fl>0*ET_Stl

_ 2C

LexUnit

S e m e D e * :

DeatedDate

Crc4edBy

Q L e n m 4_Re

QF ( o n e _Rd

eLocked

SemantJcTyps

Frame

N ar

Defintìon

CreatedDate

CreotedBy

bLocked

Image

Symbofcflep

~

^

•

4


3/11

REPORTS ON

LEXICOGRAPHICAL AM)

LEXICOLOGICAL PROJECTS

including

th e

name

and

definition,

one

fo r

frame

elements,

and

a

relation between

them

(the

line

marked

"A"

in

Fig.

)

such

that

each

FE

is

associated

with

exactly

one

frame. Lexical

units (LUs) are represented as a

table

linking lemmas

and

frames, i.e. an LU is a Saussurean

sign

linking

form

and

meaning;

there

is

also

a

field

fo r

a

description of

th e

sense.

Lemmas,

in turn

are

composed

of

one

or

more

lexemes, and lexemes

have

one

or

more word

forms.

For example,

th e

lexeme

grill

(with

word

forms grilled,

grilling, etc.)

is

th e

only

lexeme

of

a

l emma

which

is

associated

with

tw o

quite

different frames,

Questioning and

Apply_heat.

Of particular

interest

to th e present discussion is th e relation between

lemmas

and

lexemes,

which s many o

many

the in e

marked B"

n

Fig. ),

meaning

hat

emma

an

e

comprised

of more

than

one lexeme (multiword

expressions,

MWEs), and a

lexeme

can be

associated

with

more

than

one lemma. Fo r

example,

th e lemmas

write

up,

write

down

and

write

in are

al l associated with th e Writing

frame on th e one hand,

and all

contain

th e

lexeme

write,

sharing

its

word

forms

writing,

wrote,

etc. on

th e

other.

There are

three procedures by which data

are made

part

ofthe

database: (1 )

th e

annotation

process,

hrough

hich,

or pecifically

argeted

exical

nit,

xemplary

entence

constituents

are tagged

according to their

semantic

and

syntactic

relation

to th e target;

(2 )

descriptions

of frames

repared

y

he

rame

nalyst, nd

3)

elations mong

rames

prepared by

th e

lexicon

analyst.

{Databo» T r e e SW>l

S e nte nc e

pl a c e d

f l rd in

load

q u eu e

E

¡CtoMjyf

Figure 2.

Screenshot

ofthe

Annotation Tool

361


4/11

EüRALEX 2002 PROCEEDINGS

The

Annotator, shown n Figure

2 ,

consists

of

three rames

The

top

eft one

contains

a

listing

ofthe

names

ofindividual subcorpora within

which

annotation

is

carried

out ,

derived

from

searches

o r

the

target

word (write)

n

certain

predefined

syntactic

contexts.

The to p

right

frame lists

th e sentences ofthe

currently

selected

subcorpus.

Th e lower

frame is

th e

place here

ndividual

entences

re

nnotated

y

pplying

using he

ouse

r he

keyboard)

labels

to

the annotation layers

ofthe

constituents

instantiating

particular

semantic

roles.

The

most important

annotation layers

are

the

top

three

shown

here ,

which

represent the

Frame Element

(i.e.

frame-specific semantic role) ,

th e

Grammatical Function, and the Phrase

Type oftheconstituent

being

labeled.

Fi le Da t a B a s e F r ame Othe r

Q Shoot_pro)ect i les

Q

Si lence

Q

Skill_test

D S l e e p

D

stice

Q

Smuggl i ng

Q Sociabi l i ty

Q Social_connect lon

Q Soc la i_even t

Q

State

Q

Sta tement

Q

Str tctness

Q

Suasion1

Q

Suas lon2

Q

Suas lon3

Q

SuasJon4

Q

Swarmmg_mot ion

Q

Terr i tory

QTest2

QThefl

Q Transfer

Q Translt rve_act lon

Q

Transportat ion

D

Tr ia |

QType

Q Useful_space

Q

Use_f i rearm

Q Verbs

Q Volubi l i ty

Q Weal th iness

Qwri i i J

tH

FrameElements

Lexical

Unit s)

r me

Editor

FrameName

Nt i ng

Symbolic

Rep

bnage

Cha ng e

C rea ted Da te :

2002-02-28

17:57:56 .0

C rea ted B y : i n f in i ty

Las t Modif ied

B y : in fín i ty

Frame Definition

An Author

creates

a Text (physical ,

like

a

letter,

or

spoken,

like

a

speech) which

con ta ins mean ingfu l

l lngutst lc

tokens,

a nd may ha ve

a partJcularAddressee In mind.

Th e

topic is

not

an FE ,

t hough it may

be found Inside

th e

const i tuent

of the

Text

AUTHOR] {penned} [a

letter concerning

r ac lsmTEXT)

[to

CongressADDR) .

rjhe brothers AUTHOR]

{satd}

(not

tw o

words

TEXT]

[to each other ADDR | .

Frame Eiemt(s)

Relatk>nship s)

say:COD: 1e.recite aspeechorfom

speateFN: sa y < a

verba l

formula) .

| type

i n :FN :

i npu t

by

typ ing.

type ou tFN :

type completery, express

type up:FN:

create

a

message

by typlr

| t ype:COD: wri te using a typewri ter

| u t t s r : COD:

. make asound)orsay

| wri te

down :FN:

physical ly wri te

(a

ma

|writein:FN:

nsert aname)bywriting

wri te

ou tFN :

wri te

careful ly,

ganera l t y

|wri te

upCOD

S .V wri te)

d

wri te a

fi

wr i teCOD

a

ompose (a tex to rwo i

:


5/11

REPORTS ON

LEXICOGRAPHICAL AN D

LEXICOLOGICAL

PROJECTS

Chitd Frame

Elaboration Editor

Parent Frame

fcvrWnj|'

Addresses

HonoreeOHon)

«CopyAD«

«Copy Setect«d«

A u t o U ap

il

Mappliigs

for Child

Au tho r=

ln tent lona l^_t rea te.Aaent

Components=tntenUonaNy_crea te, Componen ts

Creatad_ent l ty

=

intenMonaity_craate.

Crea tad_enf l ty

DeplcUve-actQr=tntenuona l fy_crea te,

Depícüve-actor

ins t rumen i=

i n ten f lona i ty_crea tB.

I ns t rument

Location

=

lntentionaHy_create.

Place

Man ne r =

in tenUonaHy_craate,

Manne r

••

^

Parent

RetatiO iship

Note(s)

«Add

N ew Note>

Figure

4.

Screenshot

ofthe

Frame

Inheritance

Editing

Tool

The elationships

ditor l lows he

exicon nalyst o ecord

nheritance elationships

between

he

urrent rame

nd

ther

rames.

As igure

4

hows,

ne

an

onsider

he

semantic cenario of

writing

s

ub-type

of

intentionally reating omething. The

not

necessarily

omplete)

quivalences

etween

he

emantic

oles

of

th e w o

rames

re

indicated

as

so-called

mappings:

th e

author

ofthe

Writing

frame corresponds to

th e

agent of

th e

intentional

creation frame; etc.

Collocations

and

M W E s in

a

Samp le Text

In

order

to

get

some

idea

ofthe

MWEs

present

in

typical

texts, le t

us

consider

the

following

example

ofjournalist ic prose, taken

from

CNN.conVLAW, dated 14

February

2001.

Washington

(CNN)—

Alleged White

House

gunman Robert Pickett w as arraigned

Wednesday

at a

federal

court

in

Washington

and

ordered

held

without

bond. A

federal magistrate nformed Pickett of th e

harges

gainst

im

—

assaulting

federal

officer

with

a

deadly weapon, which carries a

maximum

of te n

years

in

prison.

Th e

magistrate

et a

preliminary

hearing

or

next Tuesday and

rdered

Pickett held

without

bond.

Pickett,

who

w as

shot

in

th e

knee

by th e

Secret

Service

after

allegedly firing tw o shots outside th e

White House, used crutches to

walk

into th e

court.

He did not enter a

plea.

A

general

way

ofapplying FN

valence

information

to th e analysis

ofa

sentence

is

to

(1)

363


6/11

EURALEX

2002

PROCEEDINGS

choose

a word

(starting

from

the

highest semantically-relevant

predicate),

(2 )

determine

the

f rames

that

this

word

evokes

in

this

context,

(3 )

notice

the

semantic

roles

of

the

participants

in each such f rame, 4) match

the semantic needs associated

with

each such r ame hence

with each sense

ofthe

word) with phrases

found in the

sentence, (5) those

which permit the

most

coherent

it ,

nd

6)

egister

he emantic tructures associated

with

he

dependent

constituents as provided by

the

selected f rame.

But the

analysis cannot

s imply

proceed on the

basis

of f rame

information built on the words

of the

text

taken

one

a t

a

time.

Many word sequences

in our

text

m us t

be

identified

as

f ixed

phrases

or

tight

collocations,

the

most obvious ones being the

proper

names

White

House,

Robert

Pickett,

and

Secret

Service, others including

held

without

bond,

assaulting

afederal

officer

with

a

deadly

weapon,

preliminary

hearing,

firing

shots,

nd

enter

a

plea.

A ll

of

these phrases re parsable

an d

semantically

transparent, bu t they re

also

entrenched: held

without

bond

s

ne

of

the

tandard phrases

or eporting a decision

n n

arraignment

hearing,

assaulting

afederal

officer with

a

deadly

weapon is

a

named offense

in American

\aw,preliminary

hearing

is

a

named step in the

criminaljustice process,

andfire

and

enter

are best treated as support verbs fo r the event nouns shot and plea respectively.

Many

ofthese words evoke subframes ofa

f rame involving

s teps

in

U.S.

criminal process.

Other

phases

of

the

process deal with

bail ,

indings

of guilt

or

innocence,

sentencing,

tc.,

and

various bortions

or

alternative

outings

hrough

he

process

uch

s

kipping

ail,

changing

one ' s

plea,

having

ajudge

dismiss

(or

the

prosecutor

withdraw)

the

charges

against

the

defendant,

nd so

n.

The

exica l

units n complex r ame

will

simultaneously

evoke

both

he hase

of

the

rocess

which

m a ps

n to

he

grammatical

tructure

of sentences

containing

a

given

lex ica l unit ,

and

the larger

event

type

ofwhich that

phase is

a

part .

Information on

Multiword

Expressions

in

the

FrameNet

Database

Information about Multiword Expressions

is

represented in , or derivable f rom, the FrameNet

database in

a

variety

ofways.

1. Multiword

Lexical Units

Certain

l e m m a s

in

the

FN

database

were

entered

as

multiword

units

from

the

start.

Examples

ofMWEs entered as

such

wil l

include lexicalized noun-noun compounds (wheel chair ,

etc.);

verb-particle emmas trip

p,

tc.),

nd

various

inds

of

id ioms

cook someone's goose,

etc.). In

some cases lemmas

originally

treated

on their

own

were

la ter recognized as

bested

treated

as

part

ofa MWE and the analysis was changed accordingly.

2. Collocations Involving

Subcategorization

Details

In

xtracting

entences

or annotation,

e m m a s

re argeted n heir ubcategorizational

contexts,

n d

om e

of

these

nvolve

the marking of

a

constituent

by particular.

Thus,

w e

searched for the

verb

object and

the noun objection in contexts where it preceded

a

to-phrase

(nobody

objected o your

decision;

he main objection o hat proposal).

These

LU s

an

occur

alone,

where

the

entity

objected to is missing but pragmatically salient

(Iobject

How

could

there

be

any

object ion?). But whenever this FE

is

made explicit in the sentence, the

syntactic

constituent

that

expresses it

has

to

be

a

prepositional

phrase

headed

by

to . (In

this

regard, hen , verbs ik e

object

differ from the particle verbs, whose particles can never be

3 64


7/11

REPORTS

ON

LEXICOGRAPHICAL AND

LEXICOLOGICAL

PROJECTS

omitted.)

To

repare

or

he

iscovery

of

such

ollocations

as

object o, prevent f rom,

interested

in ,

fond

of,

etc.),

th e

lexicographers

identify

th e

core

FEs

in

th e

words '

valence.

That is ,

or

each

LU , we

identify those

FEs

which

are most

centrally

connected

with th e

word 's meaning,

as

distinguished from those

that are

more

peripheral. Thus,

n

a sentence

like She

objected to th e

bill

in

an

angry editorial

in

the Times, th e phrase to th e bill is more

central han n n ngry editorial n he

imes.

W e

an

hen egard LUs with

exically

specified

prepositions

in their core valence as

instances

ofmult iword

units.

3.

Noun

Compounds

with

Core

FE

as Modifier

Among

he

ossibilities

or

he mplification

of

th e rames

ssociated

with oun

s

modification

y

another

noun

navy

captain)

r

by

pertinative

2

also alled

elational)

adjective

naval

commander).

Extracting

oun

ompounds ro m

he

FrameNet

atabase,

then,

will

eek

ut either noun

equences dentified

s

multiword

argets,

r

ouns with

modifiers

that

are

labeled

with

core FEs.

4 .

Collocations across Transparent Nouns

One

of th e

amilies

of

noun

ypes

we

ave agged,

which

we

all

ransparent

nouns,

includes ouns

esignating

ypes, ggregates,

arts,

ortions, lassifiers,

nitizers,

tc.,

especially

as

they occur

as

th e

first

noun

n

an

N ofN

construction.

B y

transparent"

we

mean

that

in

this

construction

th e

first noun

is transparent with

regard to

collocational

r

selection

elations

etween

he econd

N nd he xternal ontext of

th e onstruction

[Fontenelle

20xx]

or

transparent

to

number

agreement

[Svensson

1998].

3

Examples

with

th e

relevant

collocations

underlined

follow:

(1)

n

th e

1920s,

after

th e

British

literary establishment had neglected

him

fo r

forty

years,

Machen

attracted

a

coterie

ofadmirers in

th e

United States.

(2 )

Certain

trains of

Escherichia

coli (E.

olO.

or

example

re responsible

or

causing

"Traveller's

Diarrhoea".

(3 )

He

has

pinned

a little square ofmater ia lonto

both

his

knees

so

that

when

he

drives,

th e

fabric

ofhis

best trousers

will

not ru b against th e

steering

wheel.

5 .

Collocations

with

Transparent

Nouns

Th e fact that

we

have transparent nouns labeled as such makes

it possible to produce

tables

ofthese

N -N pairs, and when

w e

do

we

will find some that are

lexically

significant:77ocfc

of

geese,pride

oflions,

swarm

ofbees,

bout

oftheflu,

case

ofhepatitis,

etc. Ofthe

Type

nouns

we will

find

many

that

are

completely

general

(type, kind,

sort), others

that

are

more

special

(variety, brand, strain, breed). Thus, th e

decision

to give special status to transparent nouns

contributes

to

th e

detection

of

M W E s

n

tw o ways:

irst,

here an

be exically

elevant

pairings etween

he

tw o

ouns

n he

onstruction, ut

secondly,

means

of collecting

linguistically

relevant collocates

can

be devised

in

which

it can

be

shown

that

th e

second N ,

not th e

first,

figures in th e

collocational relation.

(Thus,

using

examples in

section

5,

we

can

detect

th e collocations

attract

admirers,

pin

material

onto his

knees,

and

Escherichia

coli

(E.coli)

auses

raveller s Diarrhoea.)

his

s

elated

o

he

ernel

ependency

graph

extraction

exercise

discussed

below.

365


8/11

E RALEX 2002 PROCEEDINGS

6.

Event Nouns

with

Support Verbs

Among the contexts

n

which

FrameNet annotators dentify external arguments or event

nouns

is

that

ofbeing a syntactic argument ofan accompanying predicate (control structures,

but ls o

upport

verbs

of

all

ypes) . ince

upport

verbs

Akimoto 989;

Mel'cuk 995,

1996,

998]

epresent he nteresting

ase

f

objects

electing

erbs

ather han

erbs

selecting

objects,

w e

wil l

find that

there

are lexicographically

interesting

pairings of

support

verb plus noun, and it

wil l

be possible to construct such tables

as

the

following for

the

nouns

in our Statement

frame,

which show these relationships:

Support Verbs Event Nouns

make

address, admission, allegation,

announcement ,

assertion,

comment ,

complaint,

concession,

confession,

declaration, xclamation, proclamation,

remark, s tatement

give

address, exclamation, lecture

deliver

address,

lecture

issue

declaration, denial, proclamation

utter exclamation,

remark

express,

lodge,

register,

submit,

voice

complaint

face,

get

complaint

have

complaint,

revelation

Table .

Event Nouns and Associated

Support

Verbs

In

the

Statement

Frame

Table provides

everal

nteresting pieces

of

information.

t sho ws

that

make

occurs

with

the

broadest range

ofnouns

in

this

frame.

It also suggests

that

the

type

ofspeech

events

that

are

delivered

are

nes

hat

have

a

public

audience

ather

han just n nterlocutor

s

n

Addressee.

In

addition

the

table

shows ,

in an indirect

w a y ,

that

various semantic

roles of

an

event noun can be realized

as

the subject ofdifferent support verbs. Consider the fact that the

noun complaint

occurs

n

four

different

rows

n

the

table.

The verbs

in

the irst

two

of the

rows

ake

he

perspective

of

the

peaker

mak e ;

xpress,

tter,

odge,

egister,

ubmit),

whereas those

in

the last two rows

tface,

get;

have)

take the perspective of the addressee.

This

distinction is exemplified

fo r

complaint

by

the

sentences in (4-5)

and

(6-7).

(6) The woman MADE no complaint

to

the police for five

days .

(7 )

Voters

have

VOICED

complaints

at the

elections being held before

the

trials

begin nd

before Mr Papandreou has

a

chance

to

prove

his

innocence.

(4) Bernard Antoine,

general manager of the Novotel, West London,

said

he

had RECEIVED

no

complaints about charges.

(5 )

om e

ofthe things he

says

are

really

quite outrageous;

do

you ever

GET

any complaints

about

h is

language?

Inspection of support

verb

patterns cross

many emantic

r ames would

ikely

upport a

larger currently

n ly ntuition-based) eneralization

hat face nd

get lways

xpress

patients

(o r

a t

least ,

non-agents).

3 66


9/11

REPORTS

ON LEXICOGRAPHICAL AND

LEXICOLOGICAL

PROJECTS

Kernel

Dependency

Graphs KDGs)

One

of th e

side

activities

of th e

FrameNet

work

is

that of

devising

a

means of

extracting

what

w e

re alling kernel dependency graphs, by which we

mean

isplays

of

frame-

bearing

lexical

units

found in

th e

corpus together with

the

lexical

heads

of

th e

constituents

that

realize

their

core

FEs.

(in

th e

case

of

phrases marked with function words , we would

want this

to include information about

the

marker

and

th e head of th e marked constituent.

This

would effectively

be

a

display of governors together

with

their dependents along with

an

indication

of

both

th e

semantic

roles

and

the

marking

of

those dependents.

Thus, or

a

sentence

like (8 )

th e top-level

K D G could

be

represented as

in

(9).

(8 ) Th e

patient

objected

strenuously

to th e diet

her

doctor

put

her

on.

(9)

o ject

actor

p tient

content diet

marker

t o ]

One

of

th e advantages

of

recognizing that

arguments

can be found

at a distance from

th e

predicates

hey

re

emantically

ependent

n,

hrough ontrol tructures of th e amiliar

kind,

upport

verbs nd

transparent

ouns,

s

hat t ecomes

ossible

o

ero

n n he

semantically correct KDGs

in

th e

data.

Thus,

from

a

sentence like (10)

it

should be possible

to

detect

a K D G

as

in

(11), centered

in th e

noun objection

that

looks

almost exactly

like th e

previous

one,

by seeing through' th e control structure around likely,

th e support verb have,

and

th e

transparent

nouns;

given

th e

meaning of

th e

sentence,

patient

and

diet

are

more

appropriate

lexical

companions

to

objection

than

kind

and

sort.

(10)

hat

kind ofpat ient is

likely to have

strong

objections

to

this

sort

ofdiet .

1 1

jection

support h ve

actor

p tient

content diet

marker

t

Th e minimal

parsing

needed

fo r finding th e head nouns

can

generally be

done

automatically.

By gnoring ll of

th e

ransparent

tructures, w e an asily ind he words

eeded

or

extracting

th e

semantically

significant

KDGs

in

a

text.

Of

particular

interest

to

our

present

point,

om e of th e KDGs w e re ow ble o

ecognize

will

urn

ut

o e mportant

collocations in

their

ow n

right. ince

special

collocations

occur

not

only between verbs and

deverbal ouns nd

heir

omplements,

ut

ls o

etween

djectives

nd he

ouns

hey

modify,

expanding

th e

search fo r

KDGs beyond complementation structures to modification

structures

allows

us to

add

a

new

class

ofcollocations.

Including

examples

from

Fontenelle

[1999,

pp .

28-29]

w e

can find, by

skipping past th e

transparent

nouns,

th e collocations

in the

left

column

in

th e phrases given

in

the

right column

of

Table

2.

Th e first tw o are relations

between

verbs and objects; th e

next

tw o are

relations

between

adjectives

an d their

semantic

heads;

th e

last

tw o

are relations

between

prepositions

and their

object

nouns.

367


10/11

EüRALEX 2002

PROCEEDINGS

Collocation

Text

lay

eggs

The

hens

laid

dozens

of

eggs.

suffered fever

suffered

a bout

of fever

sound advice

a

sound piece

of

advice

a

fine

mess afine sort of mess

on

table on this

part

of th e

table

in

closet

in

this part

of

th e

closet

Table

2.

Finding

Collocations

across

Transparent

Nouns.

Sum m ary

So m e

nformation about multiword expressions is a

par t

of

the

FrameNet database because

lexicographers

hose

to enter space-separated words.as lexical units; th e rest is derivable by

searches r eports based n he nformation entered

n to

he database y other means .

Sometimes

an

argument

(typically

a

subject)

of

a

verb

that

takes

a

frame-bearing

noun

as it s

direct

object

is necessarily dentified with

a

f rame

element

ofthat object

noun .

hile

many

such verbs

(support verbs

and

other

sorts

of

lexical

unctions n the

sense

of

Mel'cuk)

add

configurational information

of

on e

kind or another to the verba l concept (features

of aspect,

point

of view,

evaluation,

tc.),

heir main

unction

n

m a n y

cases s

to

combine with he

nominal

object to express

a

verbal meaning : all

such

pairings (verb + object noun) can count

as MWEs; those in

which

the verb is lexically selected

by the noun are entrenched MWEs

and

hould

e

isted

eparately

n he

exicon.

erbs, djectives

nd

ouns

hose

semantically

asic

omplements

re

exically

pecified

s

eing

arked

y

articular

prepositions,

an e

ounted

s MWEs

nd

isted

it h

heir

repositions.

ll oun

compounds re MWEs, ut

n ly

hose which have eanings ssigned o hem beyond

whatever semant ic structure they m ay have y compositional principles wil l need separate

entries

in

the lex icon.

n

sum, the

database resulting from

the

straightforward

exicographic

practice

created

fo r

the FrameNet

project

has

proved capable ofyielding

reliable

information

about

articipation

n

ollocational

atterns

nd

ultiword

xpressions or

he

ords

covered in the database.

Acknowledgements

W e

are grateful to the

National

Science

Foundation fo r funding the work of

the

FrameNet

project hrough wo

rants,

RI 618838 Tools

or

exicon

uilding March

997~

February

000 , nd

TFVHCI 086132

FrameNet++:

n

n-Line exical emantic

Resource

nd

ts

pplication

o

peech

nd

anguage

echnology eptember

000-

August

2 0 0 3 .

he Principal nvestigators of

FrameNet++ re Charles . Fillmore, ICSI),

Dan

Jurafsky (University

ofColorado

a t

Boulder),

Srini Narayanan (SRI Internat ionaLflCSI) ,

and

Mark Gawron (San

Diego

State

University).

References

[Akimoto 1989]

Akimoto,

Minoji 1989, A Study of Verbo-Nominal

Structures

in

English.

Shinozaki

Shorin, Tokyo.

••••

•

• et al . 1998]

Baker,

C.F.,

C.J. Fillmore & J.B.

Lowe,

1998. The Berkeley FrameNet Project,

in: COLING-ACL

198:

Proceedings

ofthe

Conference,

heldat

the

University ofMontreal,

pp.

86-90,

Association

fo r

Computational

Linguistics,

Montreal.

368


11/11

REPORTS

ON

LEXICOGRAPHICAL AND

LEXICOLOGICAL PROJECTS

pFillmore

1985] Fillmore, C.J. 1985,

Frames

and

th e

Semantics

ofUnders tanding,

in :

Quaderni

di

Semantica

VI.2

rFillmore

&

Atkins

998]

Fillmore, C.J.

B.T.S.

Atkins

998,

rameNet

nd

exicographic

Relevance, in:

Proceedings

ofthe First

International

Conference

on

Language

Resources And

Evaluation.

Granada,

Spain.

[Fillmore

Baker

001]

Fillmore, CJ.

&

C.F.

Baker

001,

Frame Semantics

fo r

ext

Understanding,

n

Proceedings

of

WordNet

nd

Other

Lexical

Resources

Workshop, eld

t

North American Association

fo r Computat ional

Linguistics,

Pittsburgh.

rFillmore

t

l. 001]

Fillmore,

C.J.,

C.

Wooters

&

C.F.

Baker 001, Building

Large

Lexical

Database

Which

Provides

Deep

Semantics,

in :

B.

Tsou

&

O.

Kwong (eds.),

Proceedings

ofthe

15th

Pacific

Asia

Conference on

Language, Information

and

Computation.

Hong Kong.

fFontenelle

1999] Fontenelle,

T.

1999,

Semantic

Resources fo r Word

Sense

Disambiguation: A

Sine

Q ua

Non?,

in :

Linguistica

e

Filologia

9.

University

degli

Studi

di

Bergamo,

Italy.

[Gildea

&

urafsky

000]

Gildea, Daniel nd

aniel urafsky.

000, Automatic

abeling

of

Semantic

Roles,

In

Proceedings

ofthe

ACL

2000,

Hong

Kong.

fMel'cuk

995]

Mel'cuk,

. 995,

Lexical

unctions,

n:

L.

anner ed.)

Lexical Function

n

Lexicography

an d N LP . Benjamins,

pVIel'cuk

1996] Mel'cuk,

I.

1996, Phrasemes and

Phraseology,

in :

M.

Everaert,

E.-J.

van

der

Linden,

A. Schenk,

R. Schreuder

(eds.),

Idioms. tructural

and

Psychological Perspectives, Lawrence

Erlbaum Associates.

N ew

Jersey.

PVIel'cuk 1998] el 'cuk,

.

1998, ollocations

nd

exical unctions,

n:

.

owie

ed.)

Phraseology. heory, Analysis andApplications.

Clarendon Press.

[Svensson

1998]

Svensson,

P.,

1998. Nu m b e r

andCountabili ty

in English Nouns .

Uppsala: Swedish

Science

Press.

Endnotes

'

The word "frame" here means "section

of a window on th e screen"

2

Pertinative

adjectives

("pertainyms"

in

WordNet terminology)

generally

are not used predicatively

and

hen

odifying ouns

enerally

unction

n

ays imilar o

odifying ouns

n

oun

compounds. (Compare linguistic [adj]

society and

linguistics

[n ]

society,

Paris [n ] connection and

French

[adj] connection.)

Occasions

ofpredicate

use

ar e

found in

special

constructions

{thisproblem

is

economic

in

nature).

3

Of course we need

to

recognize hat

not

every

instance of

an

N

ofN pattern is a t ransparent

noun

structure: many

relational

nouns

occurring as

th e

first

N

in

this

pattern

can

have

a

following

q^-phrase

as

a

complement.

here

th e

same

word

occurs

in

either

such

structure

w e

can

have

local

ambiguity,

with

cases

in

which

it

is th e

first

noun

ofan7Vq^Nphrase

that is

th e relevant collocate

ofsomething

in

its

environment: compare

ea t a number ofapples with

calculate the number ofapples.

Th e

noun

number

is a paradigmatic fellow

to

such nouns as bunch, group, cluster, collection, etc.,

in

th e

one

context, and to

quantity,

size,

weight,

height,

etc.,

in th e

other.

3 69

037_2002_v1_josef ruppenhofer, collin f. baker & charles j. fillmore_collocational information in

Documents