UNIVERSAI,ITY AND INI)IVll)UAi,ITY: 
TttI,~ INTI,~RACTION O1,' NOUN I'IIRASI,~ I)I,"I'I,~I,tMINi,~RS IN COI'UI,AR CI,AUSF, S 
.h)lm C. M:tllery 
Political Science I)eparmtcnt ~md Artificml I,tclligence I.aboratory 
Massa,chusetts Institute of Technology 
545 Technology Square.N E43-797 
Cambridge, MA 02139, USA 
Arpanet: JCMA at MI'F-MC 
Ahstract 
This paper presents ',,n iml~lcmcntcd thct~ry fnr 
quanlifying noun phr:.i.,.;cs in clausc.s ctmt:mting ct)pulnr 
verbs (e.g.. "he" ~md "bcCOIllC'). Ih(~ccctling fr()nl Icccnt 
thcorcticnl work I)y Jackcn(Ioll\[ I')X31. this c(,uputati(,tal 
theory recognizes the dependence ol" the quantification 
dccisit)n tm the dcl~nitcness, in(lclhfitcncss, or cla'~nncs.s 
hi" I)olh the subject and ()hjcct of c(apulnr verbs irt 
Fnglish. J~lckcndofl's intuition al')out Ihc qtmntil\]c:~tional 
interdependence of suhject :tnd ohjcct Itz~s hccn imported 
fn,n his broader cognitive thc()ry and rel'ormulated 
~,tltin a ct)nstraint propagation iizmtcwt)rk, I-'xtcnsit,ts 
reported here includt: the additi,m of mort: :,ctive 
determiners, the expansion of determiner categories, and 
the tre~ttmcnt of dispktccd objects. A further linding is 
th::t qu:Httilicational constraints may prop~g:~tc across 
soine clausM boundaries. The algorithm is used by the 
I~.EI.AI'US Natural I.:mguage Understanding System 
(h,ring a phase of analysis that posts constraints to 
produce zt 'constraint tree.' This phase comes ~fter 
crcz~lit)n of synt~tctic deep structure zntd bcfure ~cntcntktl 
rcl'Clcltcc ill a semantic-network ntodcl. Incorp(~,'ation ()f 
the qtmntific~tion algorithm in a larger system that parses 
~cntcnccs z,nd builds semantic models from thcnt makes 
RE/.A'FUS ~hle to acquire taxonomic and identity 
inlormatit)n from text. 
Introduction 
the qtmntilicz, tion of noun phrases. 
determining their tmiversulity or individuality, ix critical 
It~r the ~,utomatic acquisition of taxonomic and ~dcntity 
intbrm:ttion from natural language sentences. Automatic 
:~cquisition can convert ordinary texts into sources of 
tzL,~onomic and identity information lior use by learning 
and reasoning programs in artificial intelligence. Such 
information can also find use in efforts to develop 
setectiun restrictions from Icxical sources. Of course, 
proper qu~mtific:~tion of noun phrases also plays a key 
role in computer programs that cnde~wor to understand 
natural language. 
The theory for computing the qtmntilicational 
status ()1" noun phr:~scs fi)r tile czt.',e ,fl" copular verbs (e.g.. 
"he' and "become') wzus inspired hy recent tho)rcticzd 
w~rk ,fl' .I;~ckt'nthflT \[1~)X31 . .l~mkcn(ltdl' n,~lcd that 
qtl;llltiliC:.lliOn ,d" n~ itlll i/tlr:l~t:s li)r cl)ptll:lr ,,orbs dcpcn(.I.'.; 
jointly on Ih¢ dole0 ~mnct~, ,fl' I-~tlt Ihc ~tfl~jut:t "NI ~ ,utd the 
()hjcct NI' \[1')83, ')(1-~)1. %\].1 Ills intuiti, m h:.~s hr.:on 
rcli~rmulntcd. :~ttgtt~Cnlt:d, ~md implcntcr~tcd in the 
RFI.A'I'US N~lUlztl I.:ulgtt:tgc Undcr~t:u~tling S~tem. 2 
lllc implemented qu~mlilicztti,m thc,)ry is used 
by RFI.ATUS us it incrcmcntMly builds :t scmzmtic 
m¢)tlel. lhts method recovers clztn~ ~md identity 
inlbrmution from ~)rdinary English sentences. Alth()ugh 
the '~ctll~.mtic rll~dt:l tllUSt hE ,)CCZl~,i(mz,ll~ queried to 
re.'st)lve qtmntil\]cutitmzd :tmt)igttitic~. the tnt. tht~0 is 
primz~rily syntactic :rod does n~t rcqtfire rues, ruing. The 
c()mpttt~di(mM simplicity and hn>zzd co~ct~lge oI" the 
thc,)ry zdh)w suct.'u.,,~,l'tH tltmntilqu:~tion ,d rl,~tlll phrz|st.:b ira 
most c()l)td;~r el;ruses. W,)tk b, in prt~grcs~ to c~tcnd tile 
an~d)'xis to p~rtitivcs zmd thcrchy yicltl zl c(~rnprehcn~,i~¢ 
analysis. While Ihi'~ ~lppr()at:h tl~cs not trc~,t ~u~.h difficult 
issues such z~s belief c()ntcxt'~ ~,,ttl mctz~l~ho,ic:fl u.'qlgcs, it 
tines :~tttlress mt,~.t lilct'nl cas,:x. Since the qu:mtific:ltion 
I. I will use "object NP" to refer to what ix I'ruquently called a 
"'predicate objecL" 
2. \[he experimental RELA'I US Naturai I.anguage 
Understanding System represents the con,bincd efforts of 
Gav:m I)t*ffy and the tmthor. Gavzm I)ttffy ix responsible for 
the parser, the calegt)rial disaml')igu;itor \[\[)ully. 1985b\]. the 
lexicon and lexicon t, tilities. The author is respunsible lbr the 
representation system, the reference system, the component 
that nmps deep structure to semantics, the qtmntification 
system, the inversio, system, and the question-answering 
co,nponent. 
35 
theory is deployed in a natural hmguage system that 
narscs sentences and bttilds a :~cmantic model from them, 
REIATUS bccotncs, among other things, a system Ibr 
acquiring class structure inlbrmation fi'om ordinary 
English texts. 
Fig. I. Sentence Processing in RELATUS 
Syntactic Analysis 
Input: Text Stream 
Output: Surface Structure 
I)cep Structure 
Sententiai C'onstruint Posting 
Ittput: .qorface Structure 
l)ecp Structure 
Output: Sentential Constraint "Free 
Sentential Reference 
Input: 
.~ntential Constraint "~ree 
Semantic Representation 
Output: 
Sentence Men, lt'd into Scm'.lntic Representation 
T'he quantificati~m algorithm is embedded in a 
scntcntial const:aint-posting process \[l)ufl'y ;,,d Mallery, 
1%41 shown in ligure 1. Scntential ctmstraint-posting 
ClC'\]tc~ a cott.slratt;t tree that ct)rrcsponds to n)ughly what 
t~m,d'~,m,~li, mal Clamm:aians call h~gica! Jbrm. !1~,e 
c,)~iStlmnt Irct' is tl.'St2tl to pcrl~.)lnt intcrscntcntial 
~cfuruncc (merging succcs:,ive sentences into a single 
scmanlic-n,=twt~rk :n::dcl) \[Mallory, 19851. The input to 
c()nNtl'ailtt-i)osling phase is b~xh surf;.Icc struetttrc :.lnd 
dccl" StlUCltlrc cantlniczdizcd hy a It'an.,,Ibrmatit)nut pal:ser 
\[I)uffy, 19:Gal. In a dcplh-first, h~ttont tip walk of the 
dccp structure, omstrainus describing grammatical 
r,;latio.:, arc posted on otto-terminal paise-graph n(~d~. 3 
When verbs in major clauses (Le.. clat;scs ()thcr than 
relative cl:m.',cs or clau:;al a(ljuncts) arc reached, they 
,~upcrvise the quantilication of noun phrases they 
ctlmmand. If these verbs arc copular vcrhs, the copular 
interct)nsu'ainl algt)tith.,n is applied. In other cases, 
a,t)ther experimental algorithm pcrli)rms qttantification 
by drawing on logical rclatitms I'mm surface structure. 
The result of this process in the sentcntial constraint tree. 
It is a hierarchical description ~1" gramnmticz, I and logical 
rclati(ms that is suitable input lbr the reference s~stcm. 
By sequentially referencing the sentences o1" a text, a 
semantic mo(lcl of the text in incrementally constructed. 
The ('opular Inlerconstraint ,klgorithm 
Within a constrainl-p~sting framework, the 
basic task of NP qtmntilicatitm is to) decide whether to 
post a conslraint marking the NI.' as an individual or a 
universal. Y, incc the ta.nk ill','~d',¢N kn~wing the spccilic 
subject and ,)bjcct o1" a l~pular vcTh. it is delegated to a 
higher ctmstitucnt, the vcrh. 4 lhis dclcgatmn is 
motiwlted h), tile prin('ip/e ,lhwtzl ~h'('ision-,:m(ing which 
holds that dct:isions shtmld I)c located where all required 
inlilrnlali(m is htllh avadahlc and i~r~xHnalc. In Ilti~ case. 
only Ihc VCl'h kll()WX Ihc ~tlcntllicn ~1" hlflh Illfllll phrases 
duc it)the hic~archr¢al ',tructtt0'c ~dgrammat~cal Iclalions. 
"l'htm, whcll ;.I VClh iX~nlS its own iut'~lCllli;ll Ct)li511';lilllS, it 
also tlirccts the tttmntilicati(m o1 NPs that it d~,ninates 
(e.g., ils subject and ~lhject). 
This pmccthtrc ~,,~,s rcl~rntuiatcd m a ¢~mstraint 
prtqmgat:t,n \[Waltz. 19751 I'ramcv.ta k ht:cut+sc I'caturcs t)f 
a single ctlnslilucnt cannt)l he tlcturmincd JntlcpcntlctHly 
o1 t~thcr tC~llSllltlCltlS ill the r,cnlcntial dcfivatitm. Since 
tlUantilicati~)ll:.ll c~ m~,traint pr~ ~pagatc.,, in t~ ~th di rccti< ms. 
this i)llX.cNs is Zl t',pe ill C'~#t.~lillH'#ll "tllt'r~'t~/lSlrttittl. 
I-oltunalcl~, lhc p~ssihl¢ slates ~fl" "qlN arc ,rely lt)rctx 
tlclinilc, intlclinitc. :rod class. I~cuatmc Ih¢ mlrnhcr of 
I':t~nsihl¢ NP statc~ is small (3) :rod the nuiHhcr of 
variahh:s is also ymall (2), a simple tahlc-lo~)kup 
alg,)r~thm "c(m)pdcs' s,bjcct and ohjcct qtmniil~caLions 
for al! p~)ssihle conliguration.s of NP dclinitcnc.ss. 5 
3. At present, the REI.AIUS system builds scntcntial 
ccmstraints using the cammical grammatical relations ()1" the 
sentence, tile quantilicz~titm status of n~xm phr;LseS, and the 
truth valuc.'s of verbs. Work is in pn~grc.ss to incorporate 
|empnr:d constraints on verbs, temporal adjectives, and various 
types of context markers. 
4. The RELATUS parser u~.'s non-standard parse graphs. A 
'kern" corresponds to a clause while a "verbal" is something like 
a verb phrase except that the kern tells it what its subject. 
object, and m(~lificrs are at constraint-posting time. For 
further "details. see Du fl'y \[1985a1. 
36 
Fig. 2. Determiner Categories 
h 
Determiner/Parameters NP Classification 
"\['he Definite 
• Fh is/rh ese De \[i n ite 
"1 "h a t/T'll ose De fin i te 
No det & singular proper noun. Delinite 
No det & possessively modified. Definite 
A Indefinite 
An Indefinite 
Another htdefinite 
Some Indefinite 
No det & phlral. Indefinite 
All Class 
Any Class 
Every Class 
No Class 
llle actual task of determining the 
quantificational status of the NPs dccompt)scs into three 
steps. 
(1) The dcfiniteness of the noun phr,'kses is 
,ascertained by examining the determiners and 
several othcr parameters. The algorithm is 
summarized by figure 2. Another algorithm 
dcscrihed by Iigure 3 is used Ibr determinerless 
plural NPs. 
(2) The quantificational slatus of the subject and 
object is (Ictermined by I(~)king each case tip in 
the table depicted hy Iigt, re 4. Putentially 
ambiguous cases (marked with an ~tstcrisk) may 
require referencing the noun phrase in the 
semantic model to rcs~lve the ambiguity. 
Example scntences li)r Ihc ca.~:es in figure 4 are 
I'Oulld in Iigure 5. 
(3) The vcrb-phra.sc node informs each NP of its 
quantilicatit)n (the results of stcp 2), and they in 
5. The conversion of constnunt propagation into a table 
loq~kup approach is po~.sible in this special ~,,asc because there 
are only two variables, the suhjcct and the objccL In the 
general ~:a:~e, the sb, e of toe table is exponetitial in the number 
of ,/ariables. 
turn post corresponding constraints on 
themselves. 
Fig. 3. Categories for Determiner-less NPs 
Characteristic of NP NP Classification 
Singular Proper Noun Definite 
Plural Indefinite 
Possessively Modified l)elinite 
Animate Pronouns Dclinite 
In his discussion, ,hlckcndoll'\[lgX3: 77-106, esp. 
8X-91.94-10hl nifty c~,tcguri/c.', determiners acc()r(ling to 
the distinction between definite and indefinite . I have 
added classm,ss It) his scheme m order tt) cope with such 
determiners as "all', "any' and "every'. While Jackendol'fs 
examples use only the determiners 'a'. "an', and 'the', I 
have Ibund intcrpretati~ms Ihr additional determiners 
which are summarized in figure 2. JackcndtflT considers 
proper nouns to be definite and the same is done here, 
except in certain cases t)l" phnal proper nouns which are 
interpreted a.s the plural indefinite (scc $21 in figure 5). 
The addition of the cla~s cate~urization calls lbr the'class 
determiners in the bottom of Iigure 2. 
The determiner. "no', is trevtcd as the negation 
of 'all.' Thus, the NP is quantified as a ckL~s and the 
copula negated. While SI0 and S18 in I\]gure 5 are valid, 
S19 is not. There are restrictions on where "no" can 
appear. It cannot modify b~)lh the subject ~md object. 
Nor can it modify the t~hjcct when the subject is 
indefinite (S19) ~r a universal ($2()). but it can when the 
suhj~:ct is tlcl\]nite (,%1~). lhc~c rcstricti~ms ~ccm 
generally valid Ibr literal cases cvcn lhough some 
idiomatic and mct,lphorical ctmsUttctions nlay vi(}late 
them. 
Vari()us casc:~ of dcturnlincr-lcss NPs are 
handled hy lhc alg,)rithm dial determines NP 
dcliniteness. Those t:ascs arc listed in figure 3. The 
indclhlllC category may hc incompletely handled hccause 
lhe thco,'y tlous not yet encompass partitives -- imlcfinite 
NPs dlat ix,rtitkm collcctiuns of individuals or universals. 
Thus, determiner-less NPs with plmal hc:;d nouns are not 
amdyzcd Ibr partitive readings. 
37' 
Fig. 4. Universe of lnterconstnfint Categorizations 
Case Sentence Determiner Classification Noun Phrase Quantification 
Subject Object Subject Object 
C1 SI. $2 Indefinite Indefinite Class Class 
C2 $3 Indefinite Class Class Class 
C3 $4 Indefinite Definite Class Class 
C4 $5, $6, $7 Definite Definite Individual* Individual* 
C5 $8. $9 Definite Indefinite Individual* Class 
C6 Sll Definite Class Individual Class+ 
C7 Sl2 Class Definite Class Class 
C8 St3 Class Class Class Class+ 
C9 $10, S14 Class Indefinite Class Class 
* Indicates the possibility ofquantificational ambiguity. 
+ Indicates that grammatical sentences must have displaced objects~ 
P:ntitJve determiners may engender two 
readings. Ihe Nl's they modify can be read as either 
collc~:tt~ms o1' individuals or universals. Some partitive 
determiners sttch tin 's()llle,' 'each,' 'nu~st', "few'. or "many' 
are tt.'.,cd It) make statements abemt subsets hi" a coil,cotton. 
With the exception of "some,' thc~,c are missing from" 
figure 2 pending research about how to determine their 
quantificatitm. 'Some' is interpreted just as an indefinite 
because o1" its high frequency. I hc detclmincrs 'all,' 
"any," and "every,' wcrc included because they refer to the 
entirety of a collection. None of the partitive 
determiners, cvcn the ~mes currently used to determine 
dassncss, will be adeqtmtcly handled until completiun of 
continuing work ,m the syntactic parse graphs and the 
interactitm characteristics of partitives. 
S{~metimcs copular verbs take adjectives in the 
object position, leaving no apparent object. Some of 
these adjectives have a displaced ohjec/ as in $2, SI I, Sl3, 
.";15 and ~16 in ligure 5. Wcrc there actually no-bject, 
the qtmntil\]cation or the subject w~tlhl bc determined in 
is~i'.'tiun (u.',ing a different algorithm). When the 
adjective has an ~bjcct. that object is u:;cd to perfimn the 
NP intcreonstraint with the subject. (Ja.,,cs C6 and C8 are 
imp~,,.,,ible ($22 and $24) unlcss the sentences have 
displact:d uhjccts (SI 1 and S13). tlowcvcr, this is not the 
cast: liar (26 where a copular verb is modal and has a 
partiuvc tletcrminer on its object. This suggests that 
pa~lilive readings of class determiners may make tht~e 
cases acceptable and that displaced objects simply make 
such a reading easier. 
Di.,pbwcd .~.hic¢'ts appear as the NPs to which 
"relative prememns'" brad in rula\[l',c clatlSCS or 
app~.',i'..ivcs. 515 prc~idcs an example of intcrc(mstraint 
acn~ss u relative clausc. There, "a phih,,~q~hcr' is the 
displaced ,~alhjcct <~1' the diN~laccd ~hjcct, ":m hmian 
StOiC.' Interestingly. "a iqfih~:,~fdlcr ' i,~ al~,,~ a dihptaced 
object v, ith rc'q~cct tu "Mary.' Recall Ihaf. COl~,Stl'.~lint 
pusdnL: pr()ccc(ls Ih~m the bl)l.t()lll t)l' the ptllhC gr.:lph up 
lira hierarchy ~ffgrammattcal rcfaltt)ns with quanlllic:ltJon 
lidhw, ing al(mg and being g~vcrncd b~, major verbs. In 
SI5. qtlanttlicatl~m intcrctmS!.lamt is lirst ~;pplJcd to a 
philusolfller' and 'an hmian sit,it" by the c¢~ptnla t,l" the 
relative clause, then, rt is applied to 'Mary' and 'a 
phih~sopl~cr' hy the major copula. Since the fits! NP 
interc~'nstramt li,~cs "a phih~snphcr' as a universal, that 
rcsttll is then carri,.'d over int{~ the mtcrc~mstraint with 
'Mary. In hoth S15 and St6. the qtmntilicati, mal 
constraint pr(~pa\[,atcs across ck, usal boundaries becm~se 
both clause share the ~me NP as an ohject and a subject. 
Ca.,,cs such as these should not lead to incemsistent 
quamil\]c:~titms. Instead. fl~e.~ slmuld aLzrcc. ,ttte:~tmg to 
the ~,,mndnc.,,s el" the algorithm. 
Jackcndc, IT \[I')83:971 a,gues that cases C4 and 
C5 in figure 4 are semantically ambiguous. This 
amhi~uity s~urns only to hold liar the determiner "the" and 
is i'es(,Ived b~- a simple rclcrence of the NP in the 
semantic representation. 6 If the ambiguous NP has no 
referent in the current discourse Ibcus \[Grosz, 1977\], then 
the NP must be a universal. If there is a referent, it is 
either a universal or an individual, and the same 
38 
Fig. 5. Sentences Exhil)iling C'apul:lr Interconstraiut 
i "C i-~c 
(S 1) A dog is not a reptile. (Ge,eric categoriz, lion \[Jackcndoff, 1983: 95\]) 
i-) c i-, c 
($2) An antelope is not similar to a Fish. 
i~C C ~C 
($3) A priest is similar to all religious figures. 
i-'c d,c 
($4) I':u'ailclism is not the panacea ofcombinatorial explosion. 
d,i d-,i 
(SS) ('lurk-Kent is the man who was given the martini by Mary. 
d " i d " i 
($6) ('l;uk-Kent is ~ul)ernlan. (hlt,,lily \[Jackcndoff. 1983: 95\]) 
d,i.d.c d.,i.d,c 
(S7) The tiger is the fiercest I~e'ast of the jungle. 
d-.i i .c 
($8) {;lark-Kent is ~ friendly super-hero. (Or,.li, ary calegori~almn \[Jackcndoff. 1983: 95\]) 
d ,Ld-.c i ,c 
($9) The tiger is a fiightening be:Int. \[.lackendoff, 1983:971 
C ~C i*c 
(S10) No m'.mmml is a reptile. 
d,i c,c 
(SII) {;eorge was similar to every prcffe.~sor in the school. 
C~'C d~C 
(S12) All syeoph'.mts ere the heart-throb of vanity. 
C'"C C ~C 
(S13) l,'very man is similar to any 5iped. 
¢ ~ C i-" C 
(S14) All men are Fallible creatures. 
d ,i i.c i÷c 
(S 15) Mary is similar to a philosopher who is close to an Ionian stoic. 
d-,i d-,i i,c 
(S 16) M:iry is .similar to the pifilosopher who is close to an Ionian stoic. 
d-,i d .i 
(S17) Clark-Kent \[.'.; Ihe man drinking the martini. \[Jackendoff. 1983:88-891 
d~i c .c 
(S18) .loe is no reptile. 
t'-," C~ 
(S19) ° A mammal is,lo reptile. 
($20)* Every mammal is no reptile. 
i.c i ,c 
($21) Ilabs :l,u ~s common ~:. fruit flies. 
tl , C "" 
($22)* The won|al! is all lawyers. 
d,i c,c 
($23) The woman ctmld bc any I.kwyer. 
C " ¢ " 
($24)* All nrmmmls t~rc eveJ~ warm-hhmded creature. 
(Dt, fi, Keness: L ~£ c) , (Quanlifiutlion: i ot c) indicates the ,ma!ysis of the NI" ,nder it. 
7"he definiteness categories." indefinite (i). definite (d). ttn,l class (c). 
?'he quanlifi¢atLon categories." mdivMual (i). class (c). 
" hldicales an ungrammalical senlenCe. 
39 
quantification should be cht)sen. Where both appear 
in the discourse lbcus, the individual reading is preferred. 
This is partictdarly important lbr C4 because the status of 
the subject is needed to predict that of object. In either 
case, both rnust have the same quantificational status. 
The analysis of NP qtmntification in COl+ular 
clauses is signilicantly smtplilicd by the Ihct that there is 
no nccd to analyze qt,antilicr scoping. "l'his li.)lk>ws fi'om 
the absence of ~, passive interpret:\[\[ion rt>r copt, lar verbs. 
rhcy are specialized in ctmvcying classificational 
inlinmation rather than exprcxsing active changes of 
~,tutc..";into there is nu agent ,rod no t)hjcct which is acted 
uptm. passive c(m:;tructJ(ms can have no meaningful 
intc!prct:~tion. Interchanging the suhjcct and the object 
either has no cl'l~:cI t)n itlcntlly stattcmcnts t)l" inverts the 
classil\]catum t,:la,thmship in t)tltcl" G.iscs. 'l'hus. the 
scln;.intJc spccialilutitm t)l" ct)ptllar ,.clhs Jn ctmvcying 
links t)f class hicr:trchic.', simplilics aspects of their 
s} nt.ictic analysis. 
Fig. 6. (.'lassificalion of ('olmlar I,inks in III"I,,VI'US 
Sulqect Re;tl ()hject l.i+,k ('l;t~ification 
Individual Umvcrsal Ordinary Classil\]cation 
Univcrs:tl Universal Genu:ic Classification 
htdiwdual Individual Identity Relation 
Either Adjective Quality 
A Glimpse At Semantics 
Since I(FI.ATIjS increment:ally constructs a 
sonuntic m(Klcl o1' the sentences it analyzes, the CUl)ular 
intcrcnnstruint algorithnl allo,s a class structure it) be 
aut(mmtically ucqtnirctl. "Ihe way in which this 
inl0rmation is represented in R|-I.AI'US explt>its the 
encoding scheme underlying English usage of Ct>l)ular 
'.,'rbs. Ihis unc¢~(ling mctht>d ulh)ws Ibur lypcs (;f linking 
ruluu(ms t() he cn'.:(~dcd iising a singtu It)ken. (i.e.. "be'). 
Ibis cnct~ding us stiunnlari/cd ili ligurc 6. Since the types 
can he dil'l~:rcntiatcd acc+~rding to the qtmntification of 
the nodes linked, the unique rcprc:;cnt;,titm of each link 
type docs not rcttuirc the introduction t)f ad hoc tokens. 
Orthnary and generic classilication arc used to construct 
the taxonomy. When two individtmls arc linked by a 'be' 
6. Such a strategy h,'m bcen ff)llowed For other types of 
ambiguous preposition and clause bindings tHirst. 1981. 1984; 
Duffy, 1985bl. 
relattt)n, idcnthy between them is represented, rdcntity 
between two universals is represented with two generic 
classificatio,ts indicating that each universal is a subset of 
the uther. For predicate adjectives, a spccml token (e.g., 
'HQ," 'HAS-QUAlITY') is used as the relation and the 
adjective as the object in order to represent a tree-place 
property \[Winston 1980, t982\]. This avoids confusion 
when a ~,Vt)l'd token has rises Both as an adjective and a 
noun. Because REI.ATUS incorporates a theory of 
#Herprelive semantics, where syntactic cant)nicalization is 
perlbrmed on input and semantic equivalence is 
dcterminetl 4)nly tim)ugh reastming twcr a s)ntactically 
canonical rt:prescntation, this cnct;tling system is 
p:,rticularly appropriate, gecat,se no p(~st-proccs.,,ing is 
ncctlctt tt) ",uhstittHc tlistinct tokens fi~r the different types 
of linking rclalitms, this cnct)tling :lls() .umplil\]cs 
quanltl\]t.alitm t)r C()l'mlar el:roses, :.llld thcrel't)rc, the 
ctmsttamt p()sting p!<)ccns in gcnc!ul. The cnct)ding 
rtlcth()(l t)011+,,' l~tltlilCS .3 '-,mall c(Hl'-,tctrtl, jnLIL':~:',c i\[I time 
li>r walking the t.rcatctl clas.,, struc'tttrc, lhtls, the 
pq)tcntial gain in cllit.Jcr+cy hy ur, ing a mare uxplictt 
cnct)tling technique ts hint gin:if+ :mtl might hc ~)l'l~,ct by 
t)l hc! I,~ctors. 
Conclusions 
lhc ~opula, huurcon,~tra,nt algoriflun prc,xCflLcd 
ill tills paper flus hccn ~llll)lL,,,ill~L,, rtd3tlS, t m l;.ll~C L.d.\L 
applicatltms ovcr th,: pa:,t +vuai. ()nee the research on 
r~;+irt+tJvu.s iis c¢)nli'Jlctcd, the :+Ll+t)l :thm will ut)~ur :at+ even 
lalger i+l(+p(+rti(+n (~l C~+l+tll~tl- vulb ca~cs. Wt!llk 11:.15 hccn 
dollc 011 c()ptllnr qt~c~tit)ns hut Ls too c,,mplcx Ibr 
disctts,;Rm hurt:, largely due t(+ pragrnuttc inturacu++ns. 
Conjum.titms havc he,.:n treated ju.,,t like t~rdimtry NPs, 
,gxo.'pt that crr,)r checking on,arc:+ that .ill ?;Ps ill 
conjutlctJt)vts agree in dcl\]nitcncss. Ihc idu:t urc()n~,traint 
prt~pagalitm has been ~:xtcndcd cxpcrhncnt:,lly to 
n~m-copular verbs using a dil'fcrcnt ptopugat:on 
alg<+rtthln. llae approach has hcen succcssl'ul thus rtr. 
Ht)wcvcr, nlt)ae research is required to analyze 
intcr:!cti~ms between '.andros qu:mtificatitm alg~>rithms 
and ':) .~l.',,cClt.~lJn lhc prt+f+ag'~tit,n characteristics or 
dil'furcnt verbs, accordint,, t(> their senses and rncanings. 
Oti.'.mtlllcr sc~ping. :dgtJrlthm intu'ractkm, and diflicrential 
pr()p:tgatit)n are some t~l" characteristics ()r general 
ctmstitucnt interconstraint that make it more dil'licult. In 
gencf:tl, propagation ,ff q,,antificational constraints, 
seems a promisin~ approach =o previously 
rccalciu'ant problems. Even sO, strong psychological 
claims rot,st await further research and exhaustive 
atmlyses acros~ languages. 
40 
Recent interest in developing lexicons to 
support computer understanding of natural language 
\[Walker and Amsler, 1985\] suggests die need for cfreetive 
methods of attgmcnting our Icxicographical knowledge 
using large corpora and unrestricted text. Selection 
restrictions are an important type of information to 
accumulate because they are needed not only to 
distinguish different senses of words but also to recognize 
metaphorical uses. Since accumuhttion of selection 
rcstrictio,s rctluires it, acquisition or taxonomic 
ialb.'mation is a priority. The coptflar intcrconstnlint 
algorithm introduced in this paper provides a basis lbr 
acquiring large taxonomics from unrestricted texts. A 
filter can be used to quickly pmnc all non-copt.lar 
sentences :t~ well as (lilfictflt copular sentences involving 
hclicl: and perhaps, time contexts. The remaining 
scntc))ccs can be parsed, quantified and represented in a 
large semantic model. 
This research would not only advance our 
knowledge of natur'd laxonomics and selection 
rc.~tricti~)ns h,t it woultl also generate empirical data 
tt:)clitl I~)r ttlose studying "dcfault logics' and stereotype 
hierarchies \[Minsky, 1975; Keil, 1979: Rotter, 1980; 
l-',rachrnan, 1982; Etherington and Rotter, 1983\]. One 
dil'l\]ct.ity with this research program is that an 
uncertainty principle is at work: The taxonomy used to 
determine selection restrictions itself depends on 
recognition of nlctaphors fllrottgh selection restrictions. 
Success in this lexicographical task will require the 
careful development of effective research strategies. 
Acknowledgments 
This paper was improved by comments rrom 
Jonathan Connetl, Gavan Daffy. M:lrgarct Fleck, Robert 
Ingria, David McAIIcster, Rick Lathrop, and David 
Waltz. This i'ese:lrch was encouraged mid supported in 
various ways by Hayward Alker, Mike Rrndy, Berthold 
Horn, Tom Knight, Marvin Minsky, Gerald Sussman, 
and Patrick Winston. Gavan Duffy's parse graphs made 
this research possible. The REI.AFUS system was 
dc:,igncd and implemented by the author and Gavan 
I)uffy. This research ~,as done at the Artificial 
Intelligence I,abonttory of the Massachttsctts Institute of 
Technology. Support for the l,aboratory's artificial 
intelligence research is pn)vidcd in part by the Advanced 
Research Projects Agency or the Department of Defense 
under Office of Naval Research contract number 
N00014-80-C-0505. Responsibility for the content, of 
course, remains with the author. 
References 
I~,rachman, Ronald .!., 11982\], "What 'lsa' Is And Isn't," 
I'roceedings ~" the b'ourth Bienaial Conference of The 
Canadian Society for Computational Studies of 
Intelligence, pp. 212-2t 
Daffy, (;avan, \[1985a\], "Bootstrapping Performance fi'om 
Competence: The Development Strategy of the 
R EI.,A'\['US Parser," Ibrthcoming. 
Duffy, (;avan, \[1985bl, "Catego*'ial Disamhiguation," 
Ibrthcoming. 
l)uffy, Gay:m, and .Iohn (,'. Mallory, 119841, "'Referential 
Determinism and Computational Efllcicncy: Posting 
Constraints from Dccp Structure" Proceedings of the 
/lmerican Ass,)chttiott for Artificial Intelligence, pp. 
101-105. 
Etheringto,, I)avid K., anti R;lymond Relier, \[19831, "On 
Inh._'ritance I-liclarchics With Exceptions, "' I'roceedings 
oJ'the Amcricat) Association fi~r Artificial Intelligence, pp. 
104-108. 
(;rosz. Barbara .1, \[1977\]. The Representation and U~e (~ 
Focus m Dialogue Understanding. Menh) Park: SRI 
International, Technical Note 151. 
llirst, (;raeme, \[1981\[, Anaphora in Natural Language 
Understanding: A Survey, Berlin: Springer-Verlag. 
l lirst, Graeme, \[198,11. "'A Semantic Process for Syntactic 
Disambiguation,'" Procecdb~gs of the American 
Association for Artificial httel/igence, pp. 148-152. 
J,'tckendoff, Ray S., \[1983\], Semantics tznd Cognition, 
Cambridge, Mass: M rr Press. 
Keil, Frank C., \[1979\[, Semantic and Conceptual 
Structure: ,,In Ontological Perspective, Cambridge, Mass.: 
Harvard University Press. 
Mallory, John C., 11985\], "Constraint-lntcrprcting 
Rcrcrcnce,'" Cambridge, Mass.: M rr Artificial 
Intelligence I.aboratory, AI Memo Nt~. ,q27, May. 1985. 
Mi.sky, Marvin. \[1975\]. "'A Framework fi)r Representing 
Knowledge," in \[Winston, 1975\]. pp. 211-277. 
reiter, Raymond, \[19801, "A Logic for Default 
Reasoning,'" Artificial Intelligence. pp. 81-132. 
41 
Waltz. David I,.. \[1975J, "'Un(Icrst.*,,Iding l.ine Drawings 
of Scenes with Shadows," in Wit:stun \[1975, 19-91\]. 
Walker, l)onald E., and Rohrrt A. Ainsler, \[1985l, "The 
U'~c oI" M~,chin¢-I~,cadahle I)ictionarit:s in Sublanguage 
A~lalysis,'" in Ralph Grishman and Richard Kittredge, 
editors. A'ubhm~uage: De.v'ription and Processing 
t.awrence Erlbaum Associates, 1985. 
Winston, I':llrick. \[1975l, editor. 771e Psychology c~ 
('omputer Vision, New York: McGraw-flill. 
Winston, Patrick I!., \[1981|1o "'l cvrning and Reasoning by 
Analogy," ('ummunicatinn.~' oj'the ACM. 23 (12). 
Winston, Patrkk I1., \[1982\]. "'1 earning New Principles 
From Precedents And Exercises,'" Artificial Intelligence, 
19(3). 
42 
