• No results found

Interleaved Semantic Interpretation in Environment based Parsing

N/A
N/A
Protected

Academic year: 2020

Share "Interleaved Semantic Interpretation in Environment based Parsing"

Copied!
7
0
0

Loading.... (view fulltext now)

Full text

(1)

Interleaved semantic interpretation in environment-based parsing

WilliamSchuler

Computer and Information Science Dept.

University of Pennsylvania

Philadelphia, PA 19103

[email protected]

Abstract

This paper extends a polynomial-time parsing

al-gorithmthat resolvesstructuralambiguityin input

sentences by calculating and comparing the

deno-tations of rival constituents, given some model of

the application environment (Schuler, 2001). The

algorithm is extended to incorporate a full set of

logicaloperators,includingquantiersand

conjunc-tions, into this calculation without increasing the

complexityof theoverall algorithmbeyond

polyno-mial time, both in terms of the length of the

in-put and the numberof entities in the environment

model.

1 Introduction

The development of speaker-independent

mixed-initiative speech interfaces,in whichusers notonly

answerquestionsbutalsoaskquestionsandgive

in-structions, is currently limited by the inadequacy

ofexistingcorpus-baseddisambiguationtechniques.

This paper explores the use of semantic and

prag-matic information, in the form of the entities and

relationsintheinterfacedapplication'srun-time

en-vironment,asanadditionalsourceofinformationto

guidedisambiguation.

Inparticular,thispaperextendsanexisting

pars-ingalgorithm that calculatesand comparesthe

de-notations of rival parse tree constituents in order

to resolve structural ambiguity in input sentences

(Schuler,2001). Thealgorithmisextendedto

incor-porateafullsetof logicaloperatorsintothis

calcu-lationsoastoimprovetheaccuracyoftheresulting

denotations{ andthereby improve theaccuracyof

parsing{ without increasing the complexity of the

overall algorithm beyond polynomialtime (bothin

termsofthe lengthof theinputand thenumberof

entitiesintheenvironmentmodel). Thisparsimony

is achieved by localizing certain kinds of semantic

relationsduringparsing,particularlythosebetween

quantiersandtheirrestrictorandbodyarguments

TheauthorwouldliketothankDavidChiang,Karin

Kip-per,andAlexanderKoller,aswellastheanonymousreviewers

forcommentsonthismaterial. Thisworkwaspartially

sup-(similartothewaydependenciesbetweenpredicate

andargumentheadwordsarelocalizedinlexicalized

formalismssuchastreeadjoining grammars),in

or-dertoavoidcalculatingexponentialhigher-order

de-notationsforexpressionslikegeneralizedquantiers.

2 Basic algorithm

Thissection describes thebasicenvironment-based

parser(Schuler,2001)whichwillbeextendedin

Sec-tion3. Becauseitwill cruciallyrelyonthe

denota-tions (or interpretations) of proposed constituents

inordertoguidedisambiguation,theparserwill be

denedoncategorialgrammars(Ajdukiewicz,1935;

Bar-Hillel,1953),whose categoriesallhavewell

de-ned typesand worst-case denotations. These

cat-egoriesare drawn from aminimal set of symbolsC

suchthat:

NP2CandS2C;

if;Æ2C then=Æ2CandnÆ2C:

Intuitively,thecategoryNPdescribesanounphrase

and the category S describes a sentence, and the

complexcategories=ÆandnÆdescribe`alacking

a Æ to the right' and `a lacking a Æ to the left'

respectively;so forexampleSnNPwoulddescribea

declarativeverbphraselackinganNPsubjecttoits

leftintheinput.

ThetypeTandworst-case(mostgeneral)

denota-tionW ofeachpossiblecategoryaredenedbelow,

givenasetofentitiesE asanenvironment:

T(S)=t:truthvalue W(S)=fTRUE;FALSEg

T(NP)=e:entity W(NP)=E

T(=Æ )=hT(Æ);T()i W(=Æ)=W(Æ)W()

T(nÆ )=hT(Æ);T()i W(nÆ)=W(Æ)W()

The denotation D of any proposed constituent is

constrained to be a subset of the worst-case

deno-tation W of the constituent's category; so a

con-stituent of category NP would denote a set of

en-tities, fe

1 ;e

2

;:::g, and a constituent of category

SnNP would denote a set of entitytruth value

pairs,fhe

1

;TRUEi;he

2

(2)

O(jEj )dierentelements,wherevisavalency

mea-sureofthenumberofNP symbolsoccurringwithin

theconstituent'scategory.

This paper will use the following denition of a

categorialgrammar(CG):

Denition A categorial grammar G is a formal

grammar(N;;P)suchthat:

isaniteset ofwordsw;

P isaniteset ofproductionscontaining:

!w forallw2,with2C,

!=Æ Æ foreveryrule=Æ!::: inP,

!Æ nÆ foreveryrulenÆ!::: inP,

andnothingelse;

N is thenonterminalsetfj!::: 2Pg.

and the following deductive parser, 1

which will be

extendedlatertohandlearichersetofsemantic

op-erations. Theparserisdened with:

constituentchartitems[i;j;]drawnfromI n

N,indicatingthatpositionsithroughj in

theinputcan becharacterizedbycategory;

alexicalitem[i;j;]foreveryrule!w2P if

woccursbetweenpositionsiandjintheinput;

aset ofrulesof theform:

andcanrecognizeann-lengthinputasaconstituent

ofcategory(forexample,asanS)ifitcandeduce

thechartitem[0;n;].

Thisparsercanbeimplementedinadynamic

pro-grammingalgorithm,usingtherecursivefunction:

F(x)=

are proposedconstituentsdrawn

from I

byrecordingtheresultofeveryrecursivesub-callto

F(x)in achart,then consultingthis charton

sub-sequent calls to F(x) for the same x constituent. 2

Since the indices in every rule's antecedent

con-stituents a

1 :::a

k

each cover smaller spans than

those in the consequent x, the algorithm will not

enter intoan inniterecursion;and since there are

onlyn 2

jNj dierentvaluesofx,andonly2n

dier-entrulesthatcouldproveanyconsequentx(tworule

formsfor=andn,eachwithndierentvaluesofk),

the algorithm runs in polynomial time: O(n 3

jNj).

Theresultingchartcanthenbeannotatedwithback

pointerstoproduceapolynomial-sizedsharedforest

1

FollowingShieberetal.(1995).

2

FollowingGoodman(1999).

representationofallpossiblegrammaticaltrees

(Bil-lotandLang,1989).

Traditional corpus-based parsers select preferred

treesfrom suchforestsbycalculatingViterbiscores

foreach proposed constituent, according to the

re-cursivefunction:

Thesescores canbecalculated in polynomialtime,

usingthesamedynamic programmingalgorithm as

that described for parsing. A tree can then be

se-lected,fromthetopdown,byexpandingthe

highest-scoringruleapplicationforeachconstituent.

Theenvironment-basedparserdescribedhereuses

asimilarmechanismtoselectpreferredtrees,butthe

scoresarebasedonthepresenceorabsenceof

enti-ties in the denotation(interpretation) of each

pro-posedconstituent:

wherethedenotationD(x)ofaproposedconstituent

xiscalculatedusinganotherrecursivefunction:

D(x)=

in which R (x)is a lexical relation dened for each

axiomx of category equal to some subset of 's

worst-casedenotationW(),asdenedabove. 4

The

operator onis natural(relational) join onthe elds

ofitsoperands:

AonB=fhe

the rst element of the result (corresponding the

most recently discharged argument of the head or

functorcategory):

This interleaving of semantic evaluation and

pars-ingfor the purpose of disambiguationhas much in

commonwith thatofDowdingetal.(1994),except

3

Here, the scoreis simplyequal to the numberof

non-emptyconstituentsinananalysis,butothermetricsare

pos-sible.

4

So a lexical relation for the constituent `lemon' of

category NP would contain all and only the lemons in

the environment, and a lexical relation for the

con-stituent `falling' of category SnNP would contain a

map-ping from every entity in the environment to some truth

value (TRUE if that entity is falling, FALSE otherwise):

(3)

NP[lemon]

fl

1 ;l

2 ;l

3 ;l

4 g

P:NPnNP/NP[in]

fhb

1 ;hl

1 ;l

1 ii;hm

1 ;hl

2 ;l

2 iig

NP[bin]

fb

1 ;b

2 g

P:NPnNP/NP[by]

fhm

1 ;hb

1 ;b

1 ii;hm

2 ;hb

2 ;b

2 iig

NP[machine]

fm

1 ;m

2 ;m

3 g PP:NPnNP[in]

fhl

1 ;l

1 ig

PP:NPnNP[by]

fhb

1 ;b

1 i;hb

2 ;b

2 ig NP[lemon]

fl1g

NP[bin]

fb1;b2g PP:NPnNP[in]

fhl1;l1ig fl

1 g[;

Figure1: Denotation-annotatedforestfor`lemonin binbymachine.'

that in this case, constituents are not only

seman-tically type-checked, but are also fully interpreted

eachtimetheyareproposed.

Figure 1 shows a sample denotation-annotated

forest for the phrase `the lemon in the bin by the

machine',usingthelexicalizedgrammar:

lemon, bin,machine:NP

the:NP=NP

in,by:NPnNP=NP

inwhichthedenotationofeachconstituent(theset

in each rectangle) is calculatedusing a join on the

denotationsofeachpairofconstituentsthatcombine

to produceit. In thisexample, theright-branching

tree would be preferredbecause thedenotation

re-sultingfromthecompositionattherootoftheother

treewouldbeempty.

Sincethisuseofthejoinoperationislinearonthe

sum of the cardinalities of its operands, and since

the denotations of the categories in a grammar G

areboundedincardinalitybyO(jEj v

)wherevisthe

maximum valency of the categories in G, the total

complexityofthe abovealgorithmcan beshownto

beO(n 3

jEj v

): polynomialnotonlyonthelengthof

theinputn,butalsoonthesizeoftheenvironmentE

(Schuler,2001).

3 Extended algorithm

Theabovealgorithm works well forattaching

ordi-narycomplementsandmodiers,butasasemantic

theoryitisnotsuÆcientlyexpressivetoproduce

cor-rectdenotationsinallcases.Forexample,thelexical

relations dened aboveare insuÆcient to represent

phrase`theboywithnobackpack.' 5

Asimilar

prob-lemoccurswithconjunctions;forexample,theword

`and'(usingcategoryNPnNP=NP)inthephrase`the

child wearing glasses and blue pants', also cannot

beproperlyrepresentedasalexical relation. 6

This

raises the question: how much expressivity can be

allowedinasharedsemanticinterpretationwithout

exceeding the tractable parsing complexity

neces-saryforpracticalenvironment-basedparsing?

In traditional categorial semantics (Montague,

1973;BarwiseandCooper,1981;KeenanandStavi,

1986)quantiersand noun phraseconjunctions

de-note higher-order relations: that is, relations

be-tween whole sets of entities instead of just

be-tween individuals. Under this interpretation, a

quantier like `no' would denote a set of pairs

fhA

1 ;B

1 i;hA

2 ;B

2

i;:::gwhere each A

i and B

i are

disjoint subsets of E, corresponding to an

accept-ablepair of restrictor and body sets satisfying the

quantier`no'. Unfortunately,sincethecardinalities

ofthesehigher-orderdenotationscanbeexponential

onthesizeoftheenvironmentE(thereare2 jEj

pos-siblesubsetsofE and2 2jEj

possiblecombinationsof

twosuch subsets),such anapproachwoulddestroy

thepolynomialcomplexityoftheenvironment-based

parsingalgorithm.

5

Assigning the identityrelation fhe1;e1i;he2;e2i;:::g to

thequantierwouldincorrectlyyield theset ofboyswith a

backpackasadenotationforthefullnounphrase;and

assign-ingtheconverserelation(fromeachentityintheenvironment

toeveryotherentityfhe

1 ;e

2 i;he

1 ;e

3

i;:::g)wouldincorrectly

yieldthesetofboyswithanythingthatisnotabackpack.

6

Theidentityrelationfhe

1 ;e

1 ;e

1 i;he

2 ;e

2 ;e

2

i;:::g,which

yields a correct interpretation in verb phrase conjunction,

would yield an incorrect denotation for the noun phrase

(4)

functions is restricted to a nite set (say, to some

subset of words in alexicon), it becomestractable

to store them by name rather than by denotation

(i.e. assets). Such function canthen discharge all

their rst-order arguments in a single derivational

steptoproducearst-orderresult,inordertoavoid

generatingorevaluatinganyhigher-orderpartial

re-sults. Syntactically,thiswouldbeanalogousto

com-posingaquantierwithbothanounphrase

restric-torandabodypredicate(e.g.averborverbphrase)

at the same time, to produce another rst-order

predicate (e.g. a verb phrase or sentence). Since

ageneralizedquantierfunction merelycountsand

comparesthecardinalities ofitsargumentsina

lin-eartimeoperation,thisanalysisprovidesatractable

shortcutto theexponentialcalculationsrequiredin

theconventionalanalysis.

Note that this analysis by itself does not admit

productivemodicationofquantiers(becausetheir

functionsaredrawnfromsomeniteset)orof

quan-tied nounphrases(becausetheyarenolonger

de-rivedasapartial result). Thiscausesnodisruption

totheattachmentofnon-conjunctivemodiers,

be-causeordinarysyntacticmodiersofquantier

con-stituents are seldom productive (in the sense that

their composition does not yield functions outside

someniteset),and syntacticmodiersof NP

con-stituentsusuallyonlymodifytherestrictorsetofthe

quantierratherthantheentirequantiedfunction,

and can therefore safely be taken to attach below

thequantier,totheunquantied NP.

But this is not true in cases involving

conjunc-tion. Conjoined quantiers,like`somebut notall,'

cannotalwaysbedenedusingasinglestandard

lex-ical function; and conjunctions of quantied noun

phrases, like `one orange and one lemon', cannot

beappliedto unquantiedsubconstituents

(syntac-tically, becausethis would fail to subsume the

sec-ond quantier, and semantically, because it is not

the restrictor sets which are conjoined). Keenan

and Stavi (1986)model conjunctions of quantiers

andquantiednounphrasesusinglatticeoperations

onhigher-ordersets,but aspreviouslystated,these

higher-order sets preclude tractable interleaving of

semanticinterpretationwithparsing.

Thesolutionproposedhereistotreateach

quan-tierorquantiednounphraseconjunctionasan

el-lipticalconjunctionoftwocompleterst-order

pred-icates(e.g.verbphrasesorsentences),each

subsum-ingadierentquantierandnounphraserestrictor

(in thecaseofNP conjunction),but sharingor

du-plicating a common body predicate. This analysis

requires multiple components to keep track of the

duplicated material above the conjunction, but as

longasthe numberof components is bounded, the

containing

(duplicated)

oneorange

(unduplicated)

and onelemon

(unduplicated)

Figure2: DuplicatedverbinNPconjunction.

retained. 7

Figure2showsaduplicatedverbpredicateinthe

derivation of an NP conjunction. The conjoined

constituents (the shaded regions in the gure) are

each composed oftwocomponents: oneforthe NP

itself, containing the quantier and the restrictor

predicate, and one for the verb which supplies the

bodypredicateofthequantier. Sincetheconjoined

constituents both correspond to complete

quanti-erexpressions withnounsatisedrst-order

argu-ments,theircategoriesarethat ofsimplerst-order

predicates(they are each complete verbphrases in

essence: `containingoneorange'and`containingone

lemon'). The conjunction then forms alarger

con-stituent of the same form (the unshaded outline

in the gure), with a lower component containing

theconjoinedconstituents'NPcomponents

concate-natedintheusualway,andanuppercomponentin

which the conjoined constituents' non-NP

compo-nentsareidentiedoroverlapped. Iftheduplicated

componentsdonotcoverthesamestringyield,the

conjunctiondoesnotapply.

Notethat,sincetheyareonlyappliedtoordinary

rst-orderpredicates(e.g.sentencesorverbphrases)

in thisanalysis,conjunctions cannowsafely be

as-signed the familiar truth-functional denotations in

every case. 8

Also, since the resulting constituent

hasthesamenumberofcomponentsastheconjoined

constituents, there is nothing to prevent its use as

anargumentin subsequentconjunctionoperations.

Asamplemulti-componentanalysisforquantiers

isshown below, allowing materialto be duplicated

bothtotheleftandtotherightofaconjoinedNP:

some,all,no,etc.:XnNP

q NP

q nNP

q NP

q =NP

X=NP

q NP

q nNP

q NP

q =NP

Thelexicalentryforaquantiercanbesplitinthis

7

Dahl andMcCord (1983) propose a similarduplication

mechanismtoproduce appropriatesemanticrepresentations

forNPandotherconjunctions,butfordierentreasons.

8

e.g. for the word `and': fh:::TRUE;:::TRUE;:::TRUEi;

(5)

est)of whichis notduplicated inconjunctionwhile

others may or may not be. These include a

com-ponentforthequantierNP

q =NP

(whichwill

ulti-matelyalsocontainanounphraserestrictorof

cate-goryNP

),acomponentforrestrictorPPsand

rela-tiveclausesofcategoryNP

q nNP

q

that areattached

abovethequantierandduplicated inthe

conjunc-tion,andacomponentforthebody(a verborverb

phrase or other predicate) of category XnNP

q or

X=NP

q

. The subscript q species one of a nite

set of quantiers, and the subscript indicates an

unquantiedNP.

The deductive parser presented in Section 2can

nowbeextendedbyincorporatingsequencesof

rec-ognizedandunrecognizedcomponentsinto the

con-stituent chart items. As constituents are

com-posed, components are shifted from the

unrecog-nized sequence

1

c

to the recognized sequence

hi

i, until the unrecognized

se-quenceisempty.

Theextendedparserisdened with:

chartitems of theform [i;j;;], where is

asequenceof unrecognizedcomponents, is

a sequence of recognized components ha;b;i,

and i;j;k;a;b;c areindices in theinput. Each

characterizedbythecategories

1

respectively,so

thatifthesespansareconcatenatedinwhatever

ordertheyoccurin theinputstring,theyform

a grammatical constituent of category with

unrecognizedcomponents.

alexicalitem[i;j;;hi;j;i]foreveryrule!

Tworulesto invokeleft andrightfunction

ap-plicationto anexisting component:

[i;k ;=Æ;hi;k ;=Æi][k ;j;Æ;hk ;b;Æ="i]

Tworulesto invokeleft andrightfunction

ap-plicationto afreshcomponent:

[i;k ;=Æ;hi;k ;=Æi] [k ;j;=ÆÆ;]

Tworulesto dischargeemptycomponents:

[i;j;=ÆÆ;]

[i;j;;]

[i;j;nÆÆ;]

[i;j;;]

Three rules to skip conjunctions, by adding a

ate apartial result of categoryConj 0

Æ

, and the

lattertwousethisto skiptheopposingNP):

[k ;j;Æ;]

[i;j;Conj 0

Æ ;]

[i;k ;Conj ;hi;k ;Conji]

[k ;j;Conj

Two rules to reassemble discontinuous

con-stituents(again,usingapartialresultConj 0

Æ to

reduce thenumberofrangingvariables):

[a;c;Conj;ha;c;Conji][i;j;;hc;b;Æi]

Tworulestocombine adjacentcomponents:

[i;j;;ha;c;Æ="ihc;b;"i]

[i;j;;ha;b;Æi]

[i;j;;hc;b;Æn"iha;c;"i]

[i;j;;ha;b;Æi]

Andoneruletoapplyquantierfunctions:

[i;j;;ha;b;Æ

q i]

[i;j;;ha;b;Æi]

The parsingandscoring functions remain

identi-cal to those in Section 2, but an additional k =1

casecontaining amodied projection function is

now added to the interpretation function, in order

to make the denotations of quantied constituents

dependontheirassociatedquantiers:

D(x)=

Themodiedprojectionfunction evaluatesa

quan-tier function q on some argument denotation A,

comparing the cardinality of the image of the

re-strictorsetinAwiththethecardinalityofimageof

theintersectedrestrictorandbodysets inA: 9

Thisalgorithmparsesacategorialgrammarinthe

usual way { constituentsare initially added to the

chartassinglecomponentscoveringacertain yield

in the input string (the indices of the component

arethesameastheindicesoftheconstituentitself),

and theyare combinedby concatenatingtheyields

ofsmallerconstituentstomakelargerones{untila

conjunction is encountered. Whenaconjunction is

(6)

0 1 2 3 4

Figure3: Sample derivationofconjoinedNP.

encounteredimmediatelytotheleftorrightofa

rec-ognizedconstituentconstituentx,andanother

con-stituent of thesamecategoryis foundimmediately

beyond that conjunction, the parser creates a new

constituentthathasthecombinedyieldofboth

con-stituents,butcopiesx'scomponentyield(thestring

indicesof x'soriginalcomponents)with nochange.

Thishastheeect ofcreatingtwonewconstituents

every timetwo existing constituents areconjoined:

eachwithadierentcomponentyield,butbothwith

the same(combined) constituent yield. These new

discontinuous constituents (with component yields

thatdonotexhausttheirconstituentyields)arestill

treatedasordinaryconstituentsbytheparser,which

combines them with argumentsand modiers until

all of their argument positions have been

success-fullydischarged,atwhichpointpairsof

discontinu-ousconstituentswiththesameconstituentyieldcan

bereassembledinto whole {oratleast less

discon-tinuous{constituentsagain.

A sample derivation for the verb phrase

`con-taining one orange and one lemon,' involving

con-junction of existentiallyquantiednoun phrases,is

shown in Figure3, usingthe aboveparse rulesand

thelexicalizedgrammar:

containing:SnNP

q

orange,lemon:NP

First the parser applies the skip conjunction rules

to obtain the discontinuous constituents shown

af-tersteps(1)and(2),andacomponentisdischarged

from each of the resulting constituents using the

empty component rule in steps (3) and (4). The

constituentsresultingfrom(3)and(4)arethen

com-posed with the verbconstituent for `containing' in

steps(5)and (6),usingthe leftattachment rulefor

freshcomponents. The quantiersare then applied

in steps(7) and (8), and theresultingconstituents

arereassembledusing theconjunction rulesin step

(9). The adjacent components in the constituent

resulting from step (9) are then merged using the

combinationruleinstep(10),producingacomplete

gaplessconstituentfortheentire input.

Sincetheparserrulesarexed,andthenumberof

componentsin anychartconstituentisboundedby

themaximumnumberofcomponentsin acategory

(inasmuch as the rules can only add a component

to the recognized list by subtracting one from the

unrecognizedlist),the algorithmmust run in

poly-nomial space and time on the length of the input

sentence. Sincethecardinalityofeachconstituent's

denotation is bounded by jEj v

(where E is the set

of entities in the environment and v is the

maxi-mumvalencyofanycategory),thealgorithmrunsin

worst-casepolynomialspaceonjEj;and sincethere

isnomorethan oneset composition operation

per-formedwhenaruleisapplied,andeachcomposition

(7)

ation),thealgorithmruns inworst-casepolynomial

timeonjEjaswell.

4 Evaluation

The extended parser described abovehas been

im-plemented and evaluated on a corpus of 340

spo-ken instructionsto simulated human-like agents in

acontrolled3-Denvironment(that ofchildren

run-ning alemonadestand, which wasdeemedsuitably

familiar to undergraduate student subjects). The

parser was run on the word lattice output of an

o-the-shelfspeechrecognizer(CMUSphinxII)and

theparserchartwasseededwitheveryhypothesized

word. The parser wasalso compared with the

rec-ognizerbyitself,inordertodeterminethedegreeto

which an environment-based approach could

com-plementcorpus-baseddisambiguation. Thesystems

wereevaluatedaswordrecognizers(i.e.ignoringthe

bracketsin the parseroutput) ontherst 100

sen-tencesofthecorpus(correspondingtotherstseven

of 33 subjects); the latter 240 sentences were

re-servedfortrainingtherecognizerandfordeveloping

thegrammarandsemanticlexicon.

Theaverage utterancelength wasapproximately

threeseconds(subsumingabout300framesor

posi-tionsin the parserchart), containinganaverageof

ninewords. Parsingtimeaveragedunder40seconds

persentence ona P4-1500MHz,most ofwhich was

spent in forestconstruction ratherthan denotation

calculation.

Accuracyresultsshowthattheparserwasableto

correctlyidentifyasignicantnumberofwordsthat

therecognizer missed (andvice versa), such that a

perfect synthesis of the two (choosing the correct

wordifitisrecognizedbyeithersystem)would

pro-duce anaverage of 8percentage pointsmore recall

than the recognizer by itself on successful parses,

andasmuchas19percentagepointsmoreforsome

subjects: 10

recognizer parser joint

subject prec recall fail prec recall recall

0 76 79 18 72 74 92

1 77 75 28 63 55 83

2 70 71 33 49 54 69

3 71 67 43 49 45 69

4 66 54 37 44 39 67

5 53 52 54 36 31 72

6 84 84 50 56 63 83

all 68 67 37 53 50 75

which indicates that the environment may oer a

useful additional source of information for

disam-biguation. Thoughitmaynotbepossibleto

imple-ment a perfect synthesis of the environment-based

10

Successful parses are those that result in one or more

completeanalysesoftheinput,evenifthecorrecttreeisnot

above gains can be realized, it would mark a

sig-nicantadvance.

5 Conclusion

This paper has described an extension to an

environment-basedparsingalgorithm,increasingits

semanticcoveragetoincludequantierand

conjunc-tion operations without destroying its polynomial

worst-casecomplexity. Experimentsusingan

imple-mentation of this algorithm on a corpusof spoken

instructionsindicatethat 1) theobserved

complex-ityofthealgorithmissuitableforpracticaluser

in-terfaceapplications, and 2) the ability to draw on

this kind of environment information in an

inter-faced application has the potential to greatly

im-prove recognition accuracy in speaker-independent

mixed-initiativeinterfaces.

References

KazimierzAjdukiewicz. 1935. Diesyntaktische

konnex-itat. In S. McCall, editor, Polish Logic 1920-1939,

pages 207{231. Oxford University Press. Translated

fromStudiaPhilosophica1: 1{27.

Yehoshua Bar-Hillel. 1953. A quasi-arithmetical

nota-tionforsyntacticdescription. Language,29:47{58.

Jon Barwise and Robin Cooper. 1981. Generalized

quantiersandnaturallanguage. Linguisticsand

Phi-losophy,4.

SylvieBillot andBernardLang. 1989. Thestructureof

sharedforestsinambiguousparsing. InProceedingsof

the27 th

AnnualMeeting oftheAssociation for

Com-putational Linguistics(ACL'89),pages143{151.

Veronica DahlandMichael C.McCord. 1983. Treating

coordinationinlogicgrammars. AmericanJournalof

Computational Linguistics,9(2):69{91.

John Dowding, Robert Moore, Francois Andery, and

DouglasMoran. 1994. Interleavingsyntaxand

seman-ticsinaneÆcientbottom-upparser. InProceedingsof

the32ndAnnualMeetingoftheAssociationfor

Com-putational Linguistics(ACL'94).

Joshua Goodman. 1999. Semiring parsing.

Computa-tional Linguistics,25(4):573{605.

E.KeenanandJ.Stavi. 1986. Asemantic

characteriza-tionofnaturallanguagedeterminers. Linguisticsand

Philosophy, 9:253{326.

RichardMontague.1973. Thepropertreatmentof

quan-ticationinordinaryEnglish. InJ.Hintikka,J.M.E.

Moravcsik,andP.Suppes,editors,Approachesto

Nat-uralLangauge, pages 221{242.D. Riedel, Dordrecht.

ReprintedinR.H.Thomasoned.,FormalPhilosophy,

YaleUniversityPress,1994.

William Schuler. 2001. Computational properties of

environment-baseddisambiguation. InProceedings of

the39thAnnualMeeting oftheAssociation for

Com-putational Linguistics(ACL'01),Toulouse,France.

Stuart M. Shieber, Yves Schabes, and Fernando C.N.

Pereira. 1995. Principles and implementation of

References

Related documents