The Correct Place of Lexical Semantics in Interlingual MT 
l,off LEVIN and Scrgei NIRENBURG 
Center fur Machine Translation 
Carnegie Mellon University 
Pitlsburgh, I'A 15213, U.S.A. 
1. lntroducti(m 
Inlerlingual MT has tyl)ically come to incltLde a syntac- 
tic analysis of source language (SI ,) text followed hy its 
semm~tic interpretation ~.uld representation in terms of a 
text meaning representation (TMR) scheme, an interlin- 
gu,'t. Recently two distinct views of the nattzre of lhe 
inlerlingua have become current - one based on a worhl 
model (e.g., Nirenburg et al., 1992) and another one based 
on the notion of lexieal conceptual structure (LCS) (e.g., 
Dorr, 1992). In this paper we analyze the role of LCS in 
the extraction of text memfing and argue that, thotlgh it 
cannot be considered an interlingua when used by itself, 
it etmtribtltes signilic+mtly 1o the Sl)eciiication of an ac~ 
teal interlingua. The task of ;el interlingual MT system 
btfilder is, then, to lind tt way to integrate the informalion 
provided in LCS into an ontology-molivaled text mean- 
ing representation serving as interlingua. In this paper, we 
propose a model for Ibis integration mid illustrate the pro- 
cesses and static knowledge sources involved, centrally 
including tile lexicon. 
In Section 2 wc propose a model of MT that involves 
both an LCS-based lexical semantic slruclure and a 'FMR 
that is not b:tsed on LCS. Because our lexicon lbrmal- 
ism does not represent LeSs, but semantic role names 
that serve ~us labels for LCS variables, we will use Ihe 
abhreviation SDI+S (for synlax-driven lexical semanlies, 
Nirenburg grad Levin, 1992) in reference 1o our system 
instead of LCS. We argue that TMR and SI)LS are both 
necessary and that they are distinct. This model forms the 
basis of lexical-semanlic treatment of lexls in the multi- 
lingual MTl)rojectMikrokosmos. In Seclion 3 we present 
specific exmnples as analyzed in Mikrokosmos. Wc illus- 
trate the static knowledge sources (primarily the lexicon) 
and the representations that are l)roduccd (syntax, lexical 
semantics, and TMR). The Mikrokosmos model is based 
on a Iheory of form-to-meaning corrcsl)ondence which 
relies on the concept of a society of microlheories inle- 
grated in a noLi-Slratiticational manner. We brielly sketch 
the main points of this theory in the \[inal SeCliOn of this 
paper, 
2. The model 
Traditionally, intcrlingual MT systems which employ a 
full-Iflown syntactic module (e.g., KBMT-g9 (Goodman 
• ~tnd Nircifl)mg, 1992) or KANT (Carl~onell et al., 1992)) 
use a single mapping between syntactic structure and in- 
lerlingua. In Mikrokosmos, we propose a different model, 
as illustrated in Figure 1. Lexical-cotlcel~lual SlrUCtures 
(LCSs) have been suggested its meaning represenlalions 
for n,'ilural language sentences produced in accordance 
with the semantic theory developed by I lale and Jaekend- 
o11" (e.g., Jackendoff, 1983) and used in MT-related exper- 
imenls by Dorr (l)orr, 1993). The inlerlingtmtext (or texl 
meaning representation, TMR) is a slructt~re which repre- 
sents meaning of texts in accor(laL~ce with Ihe ontology- 
oriented ;.ipl)roach to COml)tllational seitlantit:s (see Ni,cn- 
burg and Lcvin, 1992). 
It is convenient 1() sh-LlCttlre oLlr ,argtHlicLlt for Otis mo(lel 
arotmd the tmcstions below (refcrrine to labels in Figure 
1), which we will discuss one-hy-onc in lhe following 
subs 
Sl, 
Syntae|je \[.~,~A 
~t~, l l'U Ct tl t't~ 
.qL 1 
/mxieal- \[.~ 11 
Conceptual 
~ 4i-tl,t+t,tni+e 
2 
lnterlingua ~N~ C 
Text (TMR) 
TI, 3 
Lexlcal- 
(',olieeptuli| ~ l) 
~q t'l'uetu t'e ,~ 
TL 
,qynt ae tit'. 
S,h-uc t ul+e 
Figure 1. Data How in a KBMT System with 2 Semantic 
Modules. 
I. llow similm are structures I and 3? flow are they 
different? 
2. lie)v+, similar are mappings A and D'? Ih),,v are they 
(lill'crenf.~ 
3. lh)w is slrtlctlne 2 diffcrenl from slrilCltlres l and 3'? 
4. WILy are rel)tesentalions l, 2, .~llld 3 all lleCessary? 
2.1. Are LexicaI-Coneeplual Structures 
I,:mguage-Universal? 
Attempts have been made to use I+CSs as interlitlgtias for 
M'I' (notably, Dorr, 1993). The impetus Ibr such work 
is provided by obserwttions that in mlmy cases LCSs 
for translation equivalents are, in fact, identical. The 
many cases in which LCSs are not identical across lan+ 
gtmI,es pose prol)lems for this apprt)ach. Methodologi- 
cally, therefore, tile type o1' work in LCg-as-itlterlingua 
projects is tinding ways of resolving e'ich such cane, 
based on observing cross-linguistic divergences in re- 
alizing meanings. There is a danger Ihat some of the 
divergences will prove unlreatable at the LCS level and, 
alternatively, that solulions for some problems will neces- 
silale changes I() tile naltH'e of the L'epreseEIlalion which 
will make the resulting struclurc resemble tile original 
LCS in progressively smaller ways. The problematic 
cases will be those in which translation equivalenls can 
349 
have differerit lexical sem,'mtics. We will inention two 
snch cases here. 
The first problem arises in the context of a complex 
event, such as a merger of two companies, which can 
lie described by mentioning ,'my of its llarts (bids, tie- 
gotiations, etc.). This is particularly problematic when 
different langnages, by convention or for ease of expres- 
sion, refer to dill'ereut parts of the complex event. In fact, 
snch divergences exist even within one language. For ex- 
ample, you c:m go to a lneeting (directed motion), attend 
a meeting (activity), or be at a lneeting (state). Similarly, 
while in English one takes a taxi, using a mmsitive verb, 
the cor,esponding Japmlese for the stone event it takusi 
ni noru (get on, board, ride in a taxi), using an intrmlsi- 
tive verb with a gem argument. Even seemingly atomic 
events and states can be broken down into their :tspectual 
components to consist of events leading np to changes 
of slate that result in new slates. For exgmlple, lhe silu- 
ation of knowing something can be expressed in English 
using the stative verb know or in Japanese using a non- 
stalive verb siru (come to know) in its restlltative Ibznl 
site iru (Lit: have coine to know). In examples snch as 
these, there will be no direct correspondence at the level 
of lexical semantics in individual lmlguages. 
The second circumstance in which translation eqniva- 
lents have different lexical selmmtics is that an element of 
metaling tllat is expressed as :m m'gnment-taking predicale 
in one l~mguage might not be expressed as an argument- 
taking predicate in ~mother langnage. Well-knowu ex- 
amples from MT literature include like vs. germ verlir 
de vs. just, etc. However, lhis phenomenon is much 
more widespread than normally acknowledged in the MT 
literature. Things that are expressed as ln:dn or auxil- 
iary verbs in English, but are not verbs at all in Japanese 
inchlde m~my high-frequency meaning elements sucli as 
phase (begin, continue, finish), modality (mt4stlshottld, 
plan, expect, try), mid evidentiality (seem, appeal; look 
like). In fact, thesyntactic means for encoding these types 
of megming vary wildly among lmlguages, going far be- 
yond the well-known verb-adverb divergences. This is 
why ill the Mikrokosmos intcrlingua we represent snch 
elements of meaning as features or operators that scope 
over clauses mid propositions. 
2.2. Ilow is an lnterlingua Different from an SI)LS 
Oittput? 
In the cases described above in which a sitlgle event is 
described witll different lexical sem~mtics the meaning 
shared by each member in the set of paraphrases makes 
a better c,'mdidale for the interlingu~d semautic represen- 
tation thml does the lexic~d semantics; and it is lhis type 
of meaning that we are striving to extract ,'rod represent ill 
the interlingna text in Mikrokosmos. Additionally, while 
SDLS concentrales on the "who-did-what-to-whom" as- 
pect of text meaning, TMRs cont~dn additional meaning 
facets, such as ,'tspect, modality, evidentiality, speech acl, 
reference, etc. Finally, as TMRs ~u'e not based on lhe 
lexic:d semantics of one particular lmlguage, there is no 
special benetit to be accrued from the imposition of the 
requirement to preserve predicate-~u'gument stnlctures. 
2.3. Universals of Semantic Role Assignment 
It is very enticing to be able to apply principles of lexi- 
cal mapping theory cross-lingnistic:dly. Similarities that 
have been observed across languages inv01ve linkings of 
semantic rotes to syntactic positions or gramnialical fnnc- 
lions, transitivity alternations,and verb el:roses. The latter 
have been described if+ soln¢ detail for English by B. Levin 
(1¢)93) ~md others. Thns, to ltte extent flu+l the hypothesis 
of cross-linguistic equivalence holds, the descriptioli of 
similar phenomena in other l+mguages, for the pnrposes 
of M'I, becomes much simpler, it not ntterly trivial. 
l lowever, langnages, as :t rn\[e, have different transi- 
tivity altern'ltions (Mitamura 1999) and even when they 
have a similar transitivity alternalion, the classes of verbs 
to which they ;tpply may be different. See Mahmoud 
1999 for a discussion of the differences in the verbs Ih:lt 
nndergo the cansative-inchoative altern:ttion in English 
and Arabic) It is, of course, desirable to take advantage 
of universals, bill it is also necessary to have a syslem 
that is tlexible enongh to accommodate cross-linguistic 
V,+lriation. 
7.4. hliegl+allon of SI)I+S into hilerlingilal MT 
Taking a l)osition on the necessity of both SDLS and TMR 
has to be based on a general approach 1o nnraveling the 
form-nleaning correspondence. For example, to make a 
TMR for John began to read we need to identify a nnm- 
bcr of meaning elements, prilnarily Ihat something look 
place hefore the time of speech, which was the begin- 
ning plut,;e of a re<iding evenl carried ont by John. 2 I low 
do we lind tllese pieces of information? Tinle before the 
time of speecll is indicated by the mos7Jfiolog j of"began". 
The beginning phase is typically intlicnled h, xically by the 
verb begin in English. We know that it is the beginning 
phase of reading becanse the syntax module tells us thai 
to reed is the complement of begin. We know Ihat John is 
reading because John is the snbject of begin (once again, 
the sytllaelic module produced this element of informa- 
lion), whose lexical properties tell us thai John is also 
nnderstood as the subject of the complcment clause. In 
oilier words, it is the predicate argnnienl structure of be- 
gin (prodnced by the synlax-to-S DLS mapping procedure 
ill the lexicon entry for begin) Ihat tells ils where to lind 
ulany of the relevant pieces of information. 
l laving lhns served the purpose of identifying a part 
of the selnanlic dependency Io be represeuled in tile linal 
TMR (just as the liudings of other syslenl modtnles played 
their assigned roles as clnes for delermining paris of Ihe 
TMR strnctnre), Ihe predicate :u'gulnent slnctnre can then 
be disc:tided. In Ihe l%~llowing seclion we give sonic 
delailed exanlplcs of Ihe nlappings involved in prodncing 
SDLS OUllnll strnctures and TMl,ts :ts well :is relevant 
paris of lexicon erllries. 
3. Some Examples 
Examples in Figures 2, 3 and 4 contaiu a ntnnl'~er of rep- 
resentative phenomena which nnderscore the diflerences 
between SDLSs aud TMRs :is well ~ts ilhistrale how tile 
two structures co-exist in the Mikmkosmos processing 
model. In doing so, we also describe a lexicon design 
which accommodates both Stl'nctures. In all three exam- 
pies the SDLS is jilSi one of the cities for dclcrtnining a 
COlilponcnl of nlealling, and is not pre, served is01norphi- 
I\[l~cidcntally, therefore verb clilsses are nell suilable its sc- 
lnantio hierarchies fi~r ontology (Mitamura 1989). .... 
it COliN also be the I)eglllllllll~ p I se ofa i I citer I'etldlng iII- 
stesld jtlgl cite instance of reading-- Ihei'e is no way l{i tlelerlnine 
which in the abSellCe (if coillexl. 
350 
cally ill tile TMR. Tile examples also illustrate the use 
of constructions (Filhnore et al. 19gg, Filhnore and Kay 
1992) as a nnit of analysis alongside words, and show 
that treatment of MT divergences in this apl)roaeh simply 
falls out of tile general iltodel. The languages used ior 
illustration are English, RussimL and Japanese. Since the 
system is symmelrical, we do not identify which is the 
source langtlage and which is the target langnage ill each 
exmnple. 
Pot" each example, we list a TMR, which is the s;une 
for all of tile l{mguages, as well as synlaclic slruclures, 
semmltic role ~tssignments (SDLS), and lexicifl entries for 
each lmlguage. It should be appment that tile TMR is 
not necessarily isomorphic Io the SDLS of any of the 
languages, and that sentences Irom different languages 
cml correspond to the same TMR even if their syntac- 
tic ,'uld SDLS representations are not isomorphic. The 
Mikrokosmos TMR structure consists of clauses which 
roughly correspond to the "who did what to whont" com- 
ponent of meaning but also includes such components as 
speech acts, speaker altitudes, indices of the speech situ- 
alien, slylistic factors its well Its relalions (e.g., temporal 
titles) allIOtlg amy el + the above, alld other elements. 
The lexical enlries include three zones---syntax, se- 
mantic role ,qssignment, and maPlting to TMR. (The lirsl 
and third zones are discussed by Meyer et aI. 199 t .) The 
lirst zone specilies an LFG-style (Bresnan 1982) syntactic 
subcategorizalion frame of a predicate, including which 
grammatical functions (subject, object, COmlllentcnt, etc.) 
the predicate mtlst appear with and any requirements the 
predicate has of those funclions (case, syntactic calegory, 
specilic lexical items, etc.). The second zone, also in the 
spirit of LFG, specilies a mapping belween tile gralnlnal- 
ical fair,talons governed by it pfcdic.'ttc arRl the sctllatltiC 
roles it ~.tssigns. Semantic 1olo :kssignmenl is indicated by 
coindexing of a sy111actic slol and a semantic role slot. 
The semantic role munes used in lhe exantples are simply 
labels lot argument positions in lexical conceptual slrtlc- 
Itlres, which are not showtl here. The syntax iuld selnallliC 
role assignlnent Zotles serve+ the pllr\[)ose of Iocaling the 
imporlant participators in the sentence. For example, they 
might tell us thal the experiencer :u'gLnltcnt is in the SLIbjCCl 
slot with dative case, or Ihat the phrase functioning as file 
lheme argument is lound ill the object position. They arc 
also imporlalll ill capltniltg bolh Inrlgtiagc-specilic gcncr- 
alizalions about verb classes and universals of SClllttnlic 
role ilSs\[gnllle\[l\[. For these leaSOl/S, the syntax and Selllilll- 
tic role zones are entcial, and therefore ii/tlSl be inch\]deal 
even in cases in which they differ drastically from the 
TMR. 
The third zone of tile lexical entry spccilies portion 
of TMR that is associated wilh a lexical item and how 
the componcnls of the TMR corrcsltond to the compo- 
nents of the syntactic alld Semanlic role zones. We have 
chosen examples in which the TMR is not isomorphic 
to the synlactie and lexical senlantic zones, ill tnost of 
the examples, a lexical item specilies Ihat title of its cmlt- 
plemenls heads Ihe associated TMR. Ill these cases, Ihe 
syntactic head of the sentence corresponds 1o some kind 
of scope-taking operator or a simple feature-vahm pair ill 
TMR.. 
The examples, inci(lcnttdly, illustrate our treatment of 
MT divergences--situations in which It source langlmgc 
sentence and its target lmlguage translation differ sig- 
nilicanlly in synlactic structure, syntaclic category, or 
l)ret`licale-argun/ent slrttcltnc. No special mechanisms 
are needed to lreat MT divergences ill oln IllO(JCL All 
lhat is needed in order to translate a sentence involving a 
divergence are source and target language lexical enlries 
of tile sort ilhlstrated here Ihat mall dil\]erent synlaCliC 
strtlcttnes elite lhe S\[llllO TMR. "File reitreseltlillions i|lld 
nlechanisnts shown ill the lexical enhies are tllotivaled lor 
non-divergelll ex:tnlples and do nol have lo lie IllOdifled 
to deat with divergent examples. This is because source 
and larger language sentences :.Ire not normally eXl;Ccled 
1o be isomorllhic to tile "FMR or to each other. 
Another inlportanI fealuxe of ottr model is that it con- 
siders constructions to be basic lexical unils ah)ng with 
words. Following Filhnore el al., 19gg, we deline con- 
structions as (possibly, discontiguous) synlaclic structure 
or produclive synlaclic pattern whose meaning it is often 
impossible to derive solely based on the meanlings of its 
components. In other words, a COllSlrtlction is a COlIIbJll,'t- 
It(lit of a syl/laCfiC Sll'll(:lln'e :ll/d tile associaled sgln~|n\[\]c 
altd pragmatic representations which, once dereliCt`l, tie 
not have Io lie composiliortally itroduced by a 'I'MP, ex- 
Iraclof. CotlSlrllCtions are typically ways of expressing 
it ilteillling that are CttllvenliOll:ll ill tile sense thai they 
are I'rozcn, lind t/or synchronically deriwdlle from general 
prhlciplcs, even il' they once were. Note thltl il Iorlnalisllt 
Stlch its the I IPSG-IIke siglt of tile dictitlllal+y slrucltlre of 
tile ACQUILEX project can lie nlatle to SUpF, Oft such an 
idea, its lqlhnore and Kay (1992) show. 
4. l~exical Semantics in an ()verall Theory 
tff Form-Meaning Correslmmlence 
The MikrokosmcJs project is based on a theory of form- 
meaning cc, rrcspondcnce, whose underlying assumptions 
can Im statcd as follows: 
• Meanings are exlrac/ed from lexls on lho basis of 
all and any available clues (e.g., syntactic, mof\]/ho- 
logical, illltl lexical properties of an ilttelance). The 
exl.la¢liotl of i|leallir~g consists of collslrtlcling Ihe 
most platlsible, thollgh usually delbasilile, hyl)othe- 
sis that is conlpalible with tile evidence, makhlg il 
an abt`hlctive process (Hobbs, 1991). 
• TIlE processing of clues in Mikrokosmos is grouped 
lille inicrolheorics for clcnlelllS ill' Illearlillg SIICll IIS 
predicatc-afglllllCllt relaliollS, aspect, lelnporill lel;.l- 
lions, modality, evidcltliality, etc. Elicit nlicrolhcory 
spccilics the ways to COllSlrllcI "I'MRs for some :.is- 
pccl of nlcallhlg by idcnlifying the Val'it/llS SylIIacIiC, 
morphologic'd, :rod lexical clues Ior Ihal element of 
meillling ill individual lallgtlagcs. 
• Ill integrating the microtheories, Mikrokc, smos re- 
jects lhe lnu'e slratilicaliona\[ apprcmch shared by 
such otherwise diverse models its AI NLF' semantics 
(e.g., l \[irsl, 1987)or Mel'~,uk's MTM (e.g., Mel'~uk, 
19el). Knowledge from all kinds of areas coexists 
in tile stone rules for the determinalion of meaning 
units. 
• The clues (pieces of evidence) for an element of 
meaning can interact in complex ways. Cities can 
reinforce or contradict each t)thel: Coercion is pos- 
sible in sittlations ill which tile clues conIlicI. Inlcr- 
prctalion o\[ a clue can be dependent on which other 
lines are pl'cSellI. 
351 
e~ 
8 
v 
o.~ .~.~ 
a8 rj ~ O0~J • &~ ~i 
.~ 
~o 
~; # .~z ~ ~ ."-:. 
o 
°. "4~ 
~ oo~ ~ 
°" i ~L 
o 
,,4 ,,4~ 
~* ~ '~ ~ ~* "~ ,.z, ~" o ,,., 
Q~ ,?,_1 ~ ~J '° 
~o~ 
be 
'5 
~8.:~ 
~. o 
o~ 
.,q 
u~ 
~o 
.o o 
o ~ ' 
~'o.~ 
o 
"r ~ 
~ o 
352 
e4 
°~ 
o o 
db~ 
© 
~ ° 
~'~ 
• ,' ~, 
°° ¢.1 
~ ~ .~ ~'~ 
.. 0 o 
~ ~a~.~ 
.° 
• ~, \[ ~ 
;2 
r~ 
o 
, Z~ .~ 
~.~ ~? 
~ o.~ 
sN 
'-G '~ u 
I::1 1 m 
~.~ o ~ 
NS~ 
353 
= 
~.,m I=:1 
~ .. ~.~ 
N 
E 
E 
~ I~ i~ ~~ 
,1t 
• = 
~ .~. 
~ ~o-~.~ ~.~.~ 
E 
a,~ {. ~o i~ .~..~ 
~i ~ 
~ ~~ ~ • !~ ~.~ 
88 ~ 
• ~ ~ 
s2 ~.~N .~ .~ o o _ .~ 
~ 8"~ 
~.~ ~ ~.~.~ 
354 
• Mikrokosnlos is anlonablo 1o working wilh inconi- 
l)lole in|i.)rlliatiOli. If IlOl all of lho hlpul condilioiis of 
lho rules are 1)rosenl, sonlo \[inclines will slill be pos- 
sible, This properly is iniporlanl boCailSO we iilleiid 
Ill deal with real Io×ls, and we canllol tlol)t~ thai COlil- 
plote knowledge will he available. In the absence 
of spot(lie knowledge, Mikrokosinos falls back Oll 
probabilislic and statistical devices. 
• An iilil)O)-lalli lilctor ht the design of lho iliicrolheo- 
ries is iho, idelllilication of forms (above the Ioxical 
level) lhal are associalod with st)me aSl}Ot:l of moan- 
ing by convention, rathe, r than through conlposilional 
on-prt)(hlclivo rlllos. We li/llow FilhilorO el al., 1988 
in adopting lho conslruclion as a basic, unil of al)aly-- 
sis. 
lit COllChlSioll, ilOlO how the cxaniplos ill Figilres 2, 
3 ~ul(l 4 relalc', It) the al)ovo backgroinld assilniptioils o\[ 
Mikrokosnlos. The o×ainl/les ilhlslrate how SDLS is 
ilsed ~.IS ~1 SOllrUc c)l chios for various inicrolheorios, inohld- 
ing lhat of Ioxical-soniaillic dopelidenoy, aspect, nio(lalily, 
speech acts, olc. \]'lie nlajor lhiding of fills paper is lhal 
TMRs are not idonlical It) SDLS oulptit slruclilros, hill 
lhai the latlcr are slill nomess~u+y in Ihat they are essential 
li)r Ilio oxliiiclion t)f lllOaning \[roll( a le×l. \]'lie OXalll- 
pies also ilhistrale the coniillox iillOracli()li of lhe, wirii)llS 
cities (llorigllchi 1993). f;or it(sial(co, the Japanese verb 
DlOl(lll Call sigiial a reqilOsl-acli()n Sl)oech act but olily fill 
appo~us in a spccilic niorpho-synlactic t~nvirolllnOlll 0ioli- 
pasl, qiloSlion, Slleakor is subjecl, hearer is second object). 
In this onvironlnonl, olhor chios lake t)ll Sllocial lileailings. 
l:or OXalilplo negation and poiOlllialily serve ollly It) st)f toil 
tile assertiveness of Ihe ieqileSl. Convolllit)lialily is also 
ilhislratcd in the abt)vo exalt(pies. Mally of the oxaniplcs 
illuslraic coilslructiolls lhal ai+o associated will) SClnantio 
alid llragltlalic nioanhlgs I)y coiivolllioll. We leave tllt~ 
iSSilOS of ll(ill-Slralilicationalily and working ,,villi inconl- 
plelo hifornialion It)r fllltlro papers which deal priniarily 
wilh tile control slrucluro of Mikrokosmos. 
Another iml)Orlanl Ct)lllribilliOn of this paper is to silg- 
gosl a liamowork in which MT divergoliOOS arc lieu(+ 
died tlsiiig Ollly 1he liiochaniSlilS Ihal are, noodt',(I for liOll- 
divergent solllences, ()urlheory l}loclicls that divergences 
will arise bocailSO the SalllO olclli¢lll of tl/Oaliiiig ill dif- 
forolll lal)gilages will ilOl Ileeossariiy be expressed will) 
ist)tnorphic synlax, nlo)\])hl, flogy, and lexicai itch(S. Tile 
MikrokoSllli)S TMR and Iho sot of nlicrothoorios for all 
lho rolOvalll languages naturally liandlo 1he so-called tli+ 
vorgencos without any addilional lnecli;,inisnls. 
ileferences 
\[11 Ihesilan, J. 1982. The Menial llepresenialion of (;ra- 
inalnlical Ilelalions Canlbridge, MA: MIT l'rcss. 
121 Carbenell, J,, T. Milannira and F,. Nyborg. I992. \]'lie 
KANT MT Project. Ih'()ceedings ol TMI-92. Monheal. 
\[3\] l)orr, II. 1993. The Use c)I Lexical Senmntics in Inlerlill- 
gual Machine Tianslalilm, Machine "l)'anslation, 7:135- 
194. 
\[41 I)orr, II. 1992. Classil\]catiut~ t)l" tnachine h'allslalion di- 
vcrgellces at/el a pU)l)()sod sohlli(lll. (7O,il~Utationa\] Lm- 
guisUcs. 
151 (h~odnlan, K. and S. Niienhur1~ (eds.) 1992. KIIIMT-89: 
A Case Sillily in Knowledge-llased Maehhle Teans- 
lalkm. San Mateu: Mmgan Kalllmal(tl. 
\[61 Fillmore, 11., P. Kay and M.C. ()'ConJifn. 1988. I<teguhu- 
ity and ldioniaticily it( Oranlll)alical ConslitlcliOllS: the 
Case rifLer Alone. Language, 64:501-38. 
\[7\] Filtmclre, C. and I: Kay. 1992. Cmlstrucli{)n Olallllll~l.i. 
Course Materials. University {)f California at Beikeley. 
\[8\] llirst, (3. 1987. Senlantie hlterprelation and Res~lhu- 
tiull of Amlfiguity. Cambridge Univcrsfly l'rcss. 
19\] ttohbs, J. 199 I. hiteiprelatim) as Ahd)lclilm. l'u()ceedings 
of ACL-91. 
\[ 101 1 loriguchi, K. 1993. Extraction {)f t'ragm alic Inli:)rnlati~l n 
in a CALL System. NLP 111 Piojecl l~epoI\[, Carnegie 
Melhm l.lniversily. 
\[11\] Jackendoll, R. 1983. Selnantics and C(igniiion. Cam. 
bridge, MA: MIT hess. 
\[12\] Levm, B. 1993. English Verb Classes and AIlernao 
lions: A Prelilliinary Investigation.Chicago: The \[Jni 
velsily of Chicago l'ress. 
\[131 \[,cvin, L. and S. Nirenburg. 1994 (to appear). 
Ccmshuctiml-liased MT Lexicons. in M. Palmer, led.) 
\[ 141 Mahl(ulud, A.T. 1989. A (\]llmpan'aiive Study (if Middle 
alid lncholitive Allernalions in Arable and l,'Aiglish. 
l'h.I). \])isserhdion. Univelsity of l)itlsburgh. 
I151 Malsllliioto, Y. 1992. (in the Wordliolid ()f ,|al)alle,se 
Complex I~l'eilieales. Ph.I). Dissertatiun, Stanf(ml Uni 
versiiy. 
\[161 Mel'~.uk, I.A. 1981. Meaning-Text Models: A Reccnl 
'\['iend in Soviet Linguistics. 7'lie Annual Review of An. 
thrrq)ology. 
\[171 Meyer, i., 11. ()nystikevych and L. Carls(m. 1990. Lex- 
icogial~hic l'linciples and Design for Knowledge-Bascd 
Machine Transhdiml. CMU CMT Technical Repel( 90 
118. 
1181 Mitalllura, T. 1989. The llierarchical Orwulizalion (if 
P)'edielile Fl'allieS for |illerpl'elive Mallpil)14 in Natll- 
ral Languilge P)'oeessilill. Pti.D. Disseltati()n. {\]nivei'+ 
sily uf l>ittsburgh, 
1191 Nirellbmg> S,, J. farbonell, M. qim/ita and K, (J(md 
I)l<<tl(. 1992, Knowledge-Ilas~d Maehil)e Tralislalhin. 
~al( Maleo, CA: M(itgan Ktiill'lllatll). 
\[201 Nirenbuig, S. and L. Levin. 1992. Syntax-Ill(yen aild 
Ont(flogy-1)rivcn Lexical Semantics. In J. Pustejovsky 
and S. \[3ergler (eds.) Lexieal Semantics and Kll(iV!.'l- 
edge Representation. llerlin: Springer-Verlag. t'ages 
5-2O. 
355 
