AN ARCHITI~CTURI,; FOIl A UNIVFRSAI, LI,:XICON 
A Case Study on Shared Syntactic hfformatlml in Japanese, l lindi, Bengali, (;reel<, and English 
Naoyuki Nomura, Douglas A. Jones, Robert C. Bcrwicl< 
Massachusetts Irtstitute of Technology 
Artificial intelligertce I,aboratory 
nomura@ai.mit.edu 
Introduction. ltindi 
Given the prominence of the lexicon in most current (3) 
!inguistic theories (I.FG, HPSG, GB), lhe inventory of 
language particular information left in the lexicon 
teserves special attention. Constructing large 
computerized lexicons remains a difficult problem, 
milding a large array of apparently arbitrary information. 
I'his papers shows that this arbitrariness can bc 
mn,;trained more than might have been previously 
hought. In particular, arbitrariness of argument slructure, Bengali 
xord sense, and paraphrasability will be shown not only (4) 
o be constrained, hut also to be integrally rchttcd. Our 
radical) view is that wu'iation of lexical behavior across 
anguages is exactly like lexical variation within 
anguages, specifically, the difference lies in the presence 
~r absence of certain morphemes. For cxatnple, the fact 
hat Japanese has richer possibilities it\] certain verbal 
mtterns is derived solely from its morphological (;reek 
nventory. ~ Put another way, language parameters (5) 
;imply are the presence or absence of lexical material in 
he morphological componet\]t. Observed hmguagc 
~ariation patterns reflect morphological systematicity. 
I'he generative machinery for producing argutnent 
;tructure positions is fixed across languages. 
iAnguistie Motivation. 
A striking example underscoring universality of 
trgument structure is the familiar Spray/Load 2 
tlternation, shown in example (1). Despite the many 
m'l'acc differences in these data across htnguages, they 
,hare several essential properties. 
1 ) a. John loaded tile hay on the wagon. 
b. John loaded tile wagon with the hay. 
'(l\[)atlese 
2) a. 
taroo-wa teepu-o boo-ni mait,{i. 
Taro-NOM tape-ACC stick-DAT wrap-PRF 
'Tam wrapped the tape around the stick.' 
ttlroo-wa boo-o teepu-de Illaita. 
Tan>NOM stick-ACC tape-WITIl wrap-PRl; 
'Tam wrapped the stick with tim tape.' 
See Miyagawa, Uukui, and Tenny (1985) for at discussion of this 
'ffect. Also see Mat'titt ( 1975, pp 441-.455), tor 56 such morphemes+ 
;co below for additiotml discussion o1' these alternations and for an 
alternative analysis. 
See, e.g., Levin (1993) and sources cited thcre, for example, 
ackendoff (1990) and Emends (1991 ). 
a. shyam lathi--ko kagaz-sc lape..ta 
Shyam stick-ACC paper-with wral).PRl ? 
'Shyam wrapped the stick with paper.' 
b. shyam lathi-par kagaz lape-ta 
Shyam stick--on paper wrap.Pl,H: 
'Shyam wrapl)ed paper around tile stick' 
a. jodu lathi-ta kaagajh-dic murieche 
Jodu stick-l)F3" paper-with wrap.PST 
'Jodu wrapped the stick wilh the paper.' 
13. jodu lalhi Ic kaagaj Illurieche 
Jodu stick-on paper wrap.PST 
'Jodu wrapped the paper around the stick' 
a. o 7tc~vwE ~I~op'tm~r oczvo oxo 1~,{,7ovt 
o janis fortwse sane sto bagoni 
Janis I)I';T load.PST hay on.the wagon 
'Janis loaded the hay on the wagon' 
b. o Tt(xvwlg ~()p'ccocre 'to \[laTovt bte oavo 
o jatfis fortwse to bagoni lllC SallO 
Janis I)ET load.P,qT the wagon wilh hay 
'Janis loaded tile wagon with hay' 
All of these hmguages have exactly the same alternation 
type. Why? Let us locus orl tilt role of figure and 
ground 3 in lhese examples. By "alternation", we mean 
that in each language they}'~,,ure, as expressed as a direct 
object in the (a) cases, alternates wilh the fi'gure as 
expressed in an equivalent PP in the (b) cases, l)espite 
the differences in word order and case marking, all share 
the properly that the direct object is subject to a so-called 
l lolistic l~,ffect: 4 Crucially, tim (a) sentences differ from 
the (b) sentences in exactly the same way in each of 
these languages. In (lb), where John loads the wagon 
with hay, the wagon is understood to have a wl_jole load 
of hay, likewise if John smears the wall with paint, the 
:~By 'grotll/d', we mean the surface background involved in tile 
action represented by the verb. By 'figure', we mean the object 
that is brought into contact with tile ground, l:or example, in (1), 
the hay is the figure which is brought into contact with the wagon, 
in tills case, the ground. See Tahny (1978) and Emends (1991) for 
discussion of figure and ground in this connection. 
.t Note that this property is not overtly grammatically marked, as, 
say, the case of the direct (~bject is. See l,evin (1993) and the 
references there for additional discussion of the Holistic l",ffect. 
243 
wall is understood to have more paint than if John merely 
smears paint on the w~dl. Thus we may assume throughout 
that the word sense of the verbs as used in the (a) and 
(b) cases are essentially identical, s The goal of the 
remainder of this paper is to analyze and implement this 
insight in a particular representation given by both 
linguistic and computational theory, and apply it to MT. 
Basic Building Blocks: The Syntax of Word Formation 
We propose to replace idiosyncratic language p,'u'ticular 
information with a new generative component that links 
universal abstract lexical structures with the surface 
that derives words from a small set of primitives) These 
structures are assumed to be identical across languages. 
Vp 
John put pp 
the hook on the shelf 
John put the book on the shelf. 
forms of words for each language. This generative VP 
machinery is based on work by Hale and Keyser (1993) ///'~ 
and Pustejovsky (1991a). The basic architecture is shown / :</" 
in Figure 1. John % ;'P 
,~~--~--- "'.... 
Lexicon l \[~)nl lgormation: Generative Mechani.,,m \] \] 
~,--\[J Rules °fC°mp°siti°n~tl'asie l~;'i'aingl/J ,,eboo~ II II"'°cks: 
tl':"t°r' JJj 
Fig. 1. iSl~ Syntax of Word Fornmtion 
Crucially, only a restricted number of argument 
structures can be generated. The basic idea is that lexical 
X-bar structures are composed from the lexical categories 
N, A, V, and P (see fig. 2), into trees whose Spec(ifier) 
and Comp(lement) positions after movement yield the 
range of possible ,argument structures. The lexical entries 
are subject to a series of filters, as follows. 
Basic Building Blocks Rules of Composition 
Lexical Categories * X-bar Projection 
N (Noun) = entity • Move-Alpha 
A (Adjective) = state (including Head 
V (Verb) = event Movement) 
P (Preposition) = relation 
Filters (in lime and Keyser (1993)) 
HMC: Head Movement Constraint (Baker, 1988) 
ECP: Empty Category Principle (Chomsky, 1981) 
FI: Full Interpretation (Chomsky, 1986) 
UP: Unambiguous Projection (Kayne, 1984) 
PL: Predication Locality (Williams, 1980) 
To give a concrete example of the system, we derive 
the thematic properties of the denominal verb shelve 
from compositional machinery operative in the lexicon 
by composing the noun form shelf with an empty 
preposition and an empty verb to yield the form shelve. 
The structures are as shown in Fig. 2. In short, argument 
structure is produced by syntax operative in the lexicon 
s If an interlingua-based system does not constrain the number of 
word senses, it faces some serious computational problems as is 
shown in section 3. 
Fig. 3 shows the detailed schema for producing 
lexical entries for verbs. We only note briefly here a 
few important properties of the system in Fig. 3. First 
of all, the main verb V is formed by Head Movement 
of X, and Y if it exists, to V. The categorial wdue of X 
and Y are selected from the set {N,A,V,P}. For example, 
the denominal verb shelve is built as shown above iv 
One can view the work by Dor. (1993) and, previously, ethel 
attempts at lexical decomposition ranging hack through Scbank a~ 
essentially the same in spirit, bt, t without the detailed constrai)m 
provided by Figure 2. Wc regard similar proposals regardin~ 
"promotion" and "demotion" of arguments as essentially a reflectior 
of Move-:llpha. The novelty of our proposal is that it is not ad hoe 
that is, the same constraints independently justified in syntax all 
appear in lexic:d constructioll. The need for at non-arbitrary, i.e., ar 
explanatory, account of lexical argument structure should b( 
apparent. If lexical entries varied arbitrarily, we would Iogicall~ 
expect at least the following space of lexical possibilities, requMn 1 
upwards of a quarter-million diacritics. Let m be the number o 
semantic/thematic roles, such as Agent, Patient, Theme, Range 
Duration, and so on, and let n be tbe number of grammatical functions 
Then, when n=4 and m=50 (a typical number for traditional NLI 
systems) we have 251,176 differenttypes of lexical entries (Zi=0 t( 
n m/li!(m-i)! ). While some theories might propose this man 
distinctions, it seems clear that Ibis imposes a very considerabl 
learning and engineering burden. Many researchers, includin 
Makino (1991) and others, have noticed tim drawbacks of encodin 
thematic rules, but in the absence ofa clear alternative, still requir 
them for representing lexical intormation. Furthermore, it is commo 
knowledge that verbs pattern into certain equiwdence classes withi 
languages (e.g., Levin, 1993), but arbitrary verb classes woul 
imply arbitrary variatkm across languages. 
244 
Figure 2, by selecting P as X and N as Y. The deadjectival 
verb redden is built by selecting A as X and selecting 
nothing as Y. The ditransitive verb give selects V for 
:)oth X and Y, following Larson (1988). NP1, if it 
:xists, is the agent of the action, and NP2, if it exists, is 
:he affected object of the verb, tollowing Hale and Keyser 
~1993). 
(vp) 
(NPI) ((~,'~) XP 
Q~~X,Y E { N,A,V,P,~ } 
(NP2) 
((.S.)) 
Fig. 3. Schema for producing verb lexical entries. 
From the point of view of lexical representation and 
MT, the key constraint is that the entry for shelve has 
.:lements that correspond directly to the verb put and the 
:)reposition on in its representation. These elements then 
:~ecome awfilable for interpretation and for lranslation. 
We show below that this is also part of the difference 
:)etween English, Japanese, Hindi, mid Greek verbs. 
Analysis and Applications for the Universal Lexicon 
Computational Motivation. 
There have been some controversies about the merits 
md demerits of transfer-based MT and interlingua-based 
MT. Typical transfer-based MTs prepare completely 
Jifferent sets of word senses for component languages 
~o that mapping among the word senses is completely 
arbitrary, i.e., the complexity may be calculated as 
~ipartitc graph matching. We will assume an interlingua- 
9ased MT, which supposedly makes all the component 
angt, ages share common word senses or so-called 
:oncepts and thus is constrained regarding word senses. 
\[towever, interlingua-based MT still has substantial 
)roblems in making up word senses. The number of 
0vorcl senses, their well-definedness, and the p,oblem 
about linking surface words depend on excellent 
.exicographers. To give just one example here, the 
:filingual dictionary Sanseedoo (1990) lists all the 
{bllowing English translations tbr a Japanese verb kazaru 
Wecorate) 
"h~,~°7~ : 
kazaru: ornament; decorate; adorn; dress; 
embellish; exhibit; display; put articles on show; 
affect; be affected; grace a ;use fair words; write 
an inflated style; mince; not being plain or flat 
21ustering these into well-defined word senses is not an 
.-asy task; thus, it is hard to answer the word sense 
.luestion. Suppose we have a symbol to represent the 
:ore meaning of kazaru, which is shared by the English 
counterpart decorate. Since kazaru has the syntactic 
nature of a Spray/Load type alternation, the lexicon of a 
typical interlingua-based MT essentially provides the 
information described below. 
word sense : 
KAZARU-DECORATE 
syntactic information: 
Verb Alternation Type 1: 
{ \[AGENT\]NOM(/Y:),\[TI IEME\]ACC(~), 
\[MATERIAL\]WlTH(~) } 
Verb Alternation Type 2: 
{ \[AGENT\]NOM(/0::), ITHEME\]ACC(~), 
\[ GOAL\]DAT/ON/OVEI~. (}E.)} 
AGF, NT, THEME, GOAL and MATERIAL are 
thematic roles that are the key elements in the interlingua. 
NOM, ACC, WITII and DAT/ON/OVER are case- 
marking functions mapped to the surface case markers ' 
/Y;', '~r', '-Q:' and '~C'. Assuming a self-contained 
thematic role system and case-marking system, these 
markings are to be ewtluated on the corresponding 
example sentences and be decided independent of each 
other. I Iowever, the two argument structures in the above 
diagram are actually incolnpatible with each other because 
the same thematic role THEME is assigned to different 
referents: grotmd, the patient to be decorated, andfigure, 
the decoration to be attached to the patient. In effect, 
the MT system makes serious errors clue to its confusion 
of thematic roles derived from tim lexicon. 
Example Input: 
(6) 
Output: 
Taro-ga hana-o kazatta. 
Taro-NOM flower-ACC decorate-PRF 
Taro decorated the flower. 
In most contexts, the default reading of the inpnt 
sentence above should be interpreted as 'Taro decorated 
something with flowers.' This error was caused by the 
semantic clash in the lexicon. In order to avoid such 
errors, lexicographers could overwrite some thematic 
roles disregarding semantic criteria, but thus would spoil 
the interlingual \[oundalion. The remaining possible 
sohltion t%r this problem is to artificial divided the word 
sense (WS) into two syinbols: WSwith and WSon, two 
completely artifical word senses. 
Those two artil'icial word senses are essentially very 
similar to each other, if not identical, and will pose 
difficulties for lexicographers because they will have to 
put arbitrary links among word senses for similar words 
in two languages or within the same language. The two 
word senses put the two different argument structures in 
complete isolation once the analysis is completed and 
tim interlingua is fixed; the only thing the generation 
module of MT can do is accept the given word sense 
(WSwith or WSon)and generate only one argnment structure. 
This rigidity has a potential to generate sentences that 
245 
are unacceptably unnatural. 
The result for MT and lexicon construction is that the 
computational machinery will stay fixed across languages 
and thus uniformly constrain the complexity of argument 
structures eliminating most of the related arbitrariness. 
It is well known that word-for-word translations are 
not the paradigmatic case. The architecture we propose 
entails a significant improvement in isomorphic mappings 
between languages. However, the isomorphism is not at 
the level of words, but rather, at the level of morphological 
elements that enter into the lexical syntactic formation 
of words. Thus it is no accident that 'put the book on 
the shelf' is a near paraphrase of 'shelve the book', and 
it is no accident that 'put the book on the shelf' is a 
closer isomorphic map for the Japanese translations of 
'John shelved the book' shown in (1). The entailment is 
that 'shelve the book' has the same morphological material 
as 'put the book on the shelf', but the former has an 
empty preposition and an empty verb that incorporates a 
noun. 
In particular, we show how to replace thematic roles 
with the lexical syntax proposed in Hale and Keyser 
(1993) and augmented by work in Pustejovsky (1991 a)\] 
This technique yields several potential benefits: (i) 
robustness of the lexicon, (ii) greater flexibility in 
selecting more natural renditions of target language 
structures in translation, as in (7) below. Let us consider 
each of these in turn with specific examples. 
(7) a.John-wa hon-o tana-no ue-ni oi-ta. 
John-TOP book-ACe shelf-GEN upper place-AT put-PRI ~ 
John put the book on the shelf. 
b.John-wa hon-o tana-ni simatta. (more accurate) 
John-TOP book-ACC shelf-DAT pt, t.away-PRF 
John put away / stored the book on the shell'. 
if paraphrasability and translation must conform to 
the lexical syntactic structures in Figures 2 through 4, 
we have a natural method for producing a constrained 
space of possible translations, namely, the only structures 
that are allowed are those produced by thc mechanism 
outlined in Fig. 4. To highlight the relationship between 
paraphrasability and transhttability, consider the 
alternation behavior of several verbs within English, 
shown in (2). Notice in particular that different verbs 
participate in one or both halves of the alternations, or 
in neither half. As we will see, the same facts hold 
across distinct languages. 
Nonalternation: 
Not all verbs that participate in one half of the 
Spray/Load alternation participate in the other half, as 
(8) and (9) and (10) from English, Japanese, and Hindi 
illustrate. Interestingly, in these cases verbs across 
VAlong with Emonds (1992), we propose shifting a significant 
portion of the machinery out of the semantics and into syntax, in 
the |brm of syntactically enriched lexical entries. 
languages also pattern alike in terms of nonalternation ~. 
This gives additional support for our representation 
proposal. 
English 
(g) a. John covered the baby with the blanket. 
b.*John covered the blanket {over,onto,... } the baby. 
Japanese 
(9) a..k;~l~;)l'~/~,13i~g-gli~F3<,¢~o 
taroo-wa akanboo-o moofu-de oot-ta 
Taroo-TOP baby-ACC bhmket-WITtt cover-PRF 
'Taro covered the baby with the bhmket.' 
taroo-wa mool'u-o akanboo-ni oot-ta 
Taro-TOP hhmket-ACC baby-I)AT cover-Pl~,l: 
'Taro covered the blanket over the baby.' 
Hindi 
( I 0) a. jOn-ne kaprc-se bacce-ko dlmka 
Jolm-ERG cloth-WlTI 1 child-ACC cover 
'John covered the baby with a cloth' 
b. * jOn-ne kapre-ko hacce-ke-upar dhaka 
Jolm-ERG cloth-ACC baby-ON.TOP.OF cover 
Thus there are fot.tr logical possibilities for alternation 
and nonalternation, as illustrated in Fig. 4 
Type (a) 
Subject 
Type(b) 
Subject 
ttolistic 
V NP I \[with NP2\] 
Ground Figure 
llolistic 
V 
Figure 
NI' 2 \[{onto/into/... NPtl 
Ground 
Vcrb Class i. 
(a) John loaded Ihe wagon with hay. 
(b) Jolm lo.'~ded the hay ohm Ihe wagon. 
Verb Class ii. 
(a) * John poured the glass with water. 
(b) John poured lhe water into theglass. 
Verb Class iii. 
(a) John covered the baby with the blanket. 
(b) *John covered the blanket onto the baby. 
Verb Class iv. * 
(a) *John gurgled lhe glass with water. 
(b) * John gurgled the water into the glass. 
Fig. 4. The Spray-Load alternation. 
The nonalteruations arc critical clues to discoverin 
XSee Levin (1993) for an extensive survey of such phenomena. 
246 
he correct lexical representation for tim verbs and hence 
he structures that translate a more direct mapping across 
anguages. We propose that the key to the solution is 
hat in the non-alternating cases, prepositions are 
rworporated into the verb in lexical syntax just as shelve 
n Fig. 3. For exmnple, the lexical entries for pout, and 
'over contain prepositions as shown in (10). We assume 
hat the lcxical representation for the prepositions encodes 
peeifications for figure and ground, represented as F 
~nd G in (10). In essence, whatX on Y means is that X 
afigure on the ground Y. ht fact, this is tim essence 
~f what prepositions "mean", at an abstract level. Thus 
he prepositions are the primitives in our system, and 
oncepts such as figure and ground are deriwttive. This 
ntaihnent is illustrated by the linking lines in Fig. 5. 
cover(X,Y)+with(G,F) 
Fig. 5 An Explanation For The Spray-Load 
Alternation 
The lexical entry for load does not contain a preposition 
nd hence is free to alternate. Thus the reason why a 
erb does not participate in part of an alternation ix that 
: incorporates lexical material which clashes with 
otential complements. The reason that 'cover the blanket 
n the baby' is bad is that covet" already encodes the 
igure and ground relationship by incorporating with into 
:s representatoin--that is, the preposition with is frozen 
"lto the verb's representation. Adding an overt PP with 
n creates a clash in figure and ground relations. Adding 
compatible PP to a verbal structure with an incorporated 
' introduces redundancy, but is the structure ix still well- 
:~rmed. 
alternation Mismatches Across Languages. 
So far, then, we have seen only that verbs across 
iffcrent languages pattern alike. Surely the,e must be 
'ifferences or else MT efforts wonld have succeeded 
:rag ago . Fig. 6 displays all the logically possible 
~lationship between verbs across two languages. 
Types A, F, and K verbs behave identically in a 
language pair. For example, the introductory sentences 
(1)-(5) illustrated Type A correspondences, where 
L,=English and L2c {Japanese,Hindi,Bengali,Greek}. 9 
We have found (B,C,G,E,I,H) --the unshaded cells 
in the Figure 5--to be the richest source of cross-linguistic 
information, namely, the verbs that do not con'espond 
directly are the most informative regarding the nature of 
word formation, given the model that we have adopted. 
Not accidentally, they are precisely the ones that arc not 
only of particular interest for our framework, they arc 
polentially very difficult for machine translation, simply 
because, for some of them, the,'e is no way to take 
advantage of the surfitce similarity of argument structures. 
Recall that wc assume that the word sense for these 
verbs ix fixed across hmguages. Consider, then, the 
Type C alternation correspondence below, where 
L,=English and 1,2< { Japanese,Hindi,Bengali } : 
'~(21carly, type P verbs should constitute most of the verbal 
w~cabnlary, since most verbs do not participate in the Spray/Load 
Alternation, or in any given alternation, for that maner. Types 
(D,II,I,,M,N,O) do not correspond at all--- actually, we expect that 
these verbs do not exist, given the considerations regarding likely 
candidates across languages. 
'_l'ype- ._ 
l,l~,i 
Obj=Oround ,~ Pp=l%ure 
L2=ai: Obj=Ground, PP=Figure A 
L2-~.ii: *Obj=Ground, PP=Figure E 
1.2::aiii: Ol~j=Ground, *PP=Figure l 
L2=c~iv: *Obj=Gmund, *PP=Figure 
I.l~ii Ll-~.iii Ll=~iv 
*Obi=Ground Ol~j=Ground *Obj=Gmund 
PP=l;igureB ..... *lq~=Figurec . 
I1 
Fig. 6. Cross-Linguistic alternation model: 16 logical types for tile Spray-Load "alternation. 
Classification of patterns in Fig. 4 for a given verb in languages 1, l and 1,2 into tile following types A-P: 
(Shaded cells are exact cmsslingtfistic c!?\[respondences. ) 
247 
English 
(a) John decorated the wall with posters. 
(b) * John decorated the posters {over,onto .... } the wall. 
Bengali 
(a) raam ghar-e pht, l sajieche 
Ram room-ON flowers decorate 
'Ram decorated the room with flowe,'s' 
(b) ramn phul-die ghar sajieche 
Ram flowers-WITH room decorate 
'Ram decorated the room with flowers'. 
Hindi 
(b) raam phul-se ghar sajaya 
Ram flower-WITH room decorate 
'Ram decorated the room with flowers' 
Japan ese 
John-wa kabe-o posutaa-de kazat-ta. 
John-TOP walI-ACC poster-WITH 
'John decorated the walt with posters.' 
John-wa posutaa-o kabe-ni kazat-ta. 
John-TOP poster-ACC wall-DAT 
decorate-PRl v 
decorate-PRF 
representation, though it is not visible at the surface 
level. Let us consider another case for MT: 
Japanese to English: 
kono heya-wa kabe-ni hana-o kazar-oo 
this room-TOP waI1-DAT flower-ACC decorate-LET'S 
(coarsely) translated into: 
"As for this room, let's decorate the wall with flowers,' 
but with UL techniques might be more fluently translatec 
as, "Let's decorate this room by putting flowers on the 
wall." Even more broadly, we would like to suggest 
promising direction for the development of our system 
Consider the well-known difficulty of translatin~ 
Japanese -wa phrases into English. Given the ubiquit) 
of -wa phrases as compared with the relative rarity o 
English as-for phrases, we can conclude that renderin t 
-wa as as-jbr is not the best translation. 
Let us now consider cases in Japanese and Hindi fi 
which the preposition type element ix visible, and whicl 
overtly affects the alternation type. 
Alternation Type Change: 
There are additional crosslinguistic differences, whict 
may be observed in the surface form of the verbal structure 
In Japanese, one can add the verbal morpheme kake t~ 
oou. This change, from alternation type F to type H, i 
shown below. 
raam ghar-me phul sajaya (a) ~l~lI}-2)Jli/uJ)J~;~,{\[i~C}~ <~ \]co 
Ram room-WITH flowers decorate taroo-wa akanboo-o moofu-de oot-ta 
'Ram decorated the room with flowers' Taroo-TOP baby-ACC blanket-WITH cover-PP, F 
'Taro covered the baby witb the blanket.' 
As in the case of cover discussed above, the explanation 
is quite simple within our framework. For the English (b)* ~f'l~}~.(li~gJj?/~,JJi}c_N.<,?C_o 
verb decorate, there is an incorporated preposition, * tamo-wa mooft,-o akanboo-ni oot-ta 
namely with, in its lexical representation. There is a Taro-TOP bhmket-ACC baby-DAT cover-PRF 
type clash because the direct object cannot be both figure 'Taro covered the blanket over the baby.' 
and ground, in the case of "*John decorated posters on 
the wall"; see Fig. 7. 
* laroo-wa akanboo-o llloo\['tl-de, ooi-k,'lke-ta 
Taro-TOP blanket-ACC baby-DAT cover-OVIHGPRI: 
' Taro covered the blanket over the baby.' 
ii \[ 0 ~ taroo-wa moofu-o akanboo-ni I John Idec°rate( ) posters on( , Taro-TOP blanket-ACC baby-l)AT I 'Taro cove,'ed tim bkmket over the baby.' ooi-kake-ta cover-OVER-PP, I 
Fig. 7. Type clash with "decorate posters on the In Japanese, kake acids an aspect of 'trajectory' t, 
wall", tile verb sense. More precisely, as the gloss 'over 
suggests, kake rescues oou from its type clash just a 
To address one of our main points, cases like the one the preposition with does in English. ltowever, 'corm 
in Fig. 7 pose special problems for machine translation cannot be so rescued in English simply because there i 
because the translation of the word sense of English no morphological life raft. Note further that the exampl 
decorate into its Japanese counterpart kazeru does not in (a) just above also behaves as expected with ,'espec 
have a similar type clash. Thus the problem is that the to the llolistic effect. In (a) akanboo-o 'haby' is th 
English verb has a preposition frozen,into its lexical 
248 
direct object, and the baby is understood to be wholly 
covered. Sentence b' has no such effect regarding the 
baby. m 
In Hindi, one can replace dhaknaa ('cover') with 
dakh-denaa ('give cover'). This morphological change 
turns a type F alternation contrast into a type B, as 
described in Fig. 6. 
(a) jOn-ne kapre-se bacce-ko dhaka 
John-ERG cloth-WlTtl chihI-ACC cover 
'John covered the baby with a cloth' 
(b) *jOn-ne kapre-ko baccc-ke-upar dhaka 
John-ERG cloth-ACC baby-ON.TOP.OF cover 
(a)' jOn-ne kapre-se bacce-ko 
John-ER(\] cloth-WlTll child-ACC 
'John covered the baby with a clotlf 
dhak-di-yaa 
cover-give-PRF 
(b)' iOn-he kapre-ko bacce-ke-upar dhak-di-yaa 
John-ERG cloth-ACC baby-ON.TOP.OF covcr-givePl>,F 
Put briefly, our view is that variation of lexical 
t:~ehavior across languages is exactly like lexical 
wtriationwithin languages, specifically, the difference 
lies in the presence or absence of certain morphemes. 
Ontologically speaking, then, what language parameters 
are is the presence or absence of lexical material in the 
morphological component. The observed patterns in 
language variation is then reflected in morphological 
systematicity. For example, the fact that Japanese has 
richer possibilities in certain verbal patterns is derived 
from its morphological inventory. In specific, the reason 
that it is impossible for English verbs to behave like 
certain corresponding Japanese verbs is that English lacks 
an equivalent of tim Japanese aspectual morphetnes 
ts'ukusu 'exhaust', kakeru 'trajectory verb', etc) ~ But 
recall, we find that load, for example does behave 
precisely like its corresponding verbs in Japanese, llindi, 
Bengali, and Greek. in cases where verbs do m)t 
appear to behave alike, apparent differences are resolved 
hy a process of language particula," morphological 
behavior: lor example, the verbal suffixes (and prefixes) 
of Japanese such as -tsukusu 'exhaust' alter verb 
argument structure enough to bring them into 
correspondence with their former English non- 
counterparts. 
Conclusion 
We believe that our approach is applicable universally. 
Future work to be done is to complete our survey o1' the 
approximately 150 types of verbal alternations of Levin 
mlt might be tmderstood pragmatically to entail that perhaps the 
parents were wo,'ried about covering the baby too much, and wanted 
to allow the baby to breathe easily by allowing its bead, Ibr example, 
to remain uncovered. In brief, here it is the compositional belmvior 
of morphemes that yields different alternation pa!adigms. 
(1993), and augment our analysis with further ideas from 
Hale and Keyser (1993), Pustejovsky (1990, 1991 b), and 
others, and to extend the coverage to Japanese and other 
languages. Our highly constrained system should also 
provide highly desirable circumscription of 
computational lexicons. Given the universal aspects of 
our lexical representations, we also expect manageable 
applications to machine translation, along the lines that 
we have suggested. 
References 
\]hnonds, J.E. (1991). Subcategorization and Syntax-Based 
Theta-Role Assignment. Natttral lxmguage and 
Linguistic Theory. 9. 
llale, K. andS.J. Keyser (1993). On Argument Structure and the 
I+exical F, xpression of Syntactic Relations. In 77re View 
fi'om Bulding 20: F+ssays in Linguistics in Honor of 
Syh,ain Bromberget, Kenneth 1 lale and Samuel Jay 
Keyser (eds), MIT Press. 
Jackendoff, R. (1993). On the Role of Conceptual Slructure in 
Argument Selection: A Reply to lhnonds. Natural 
Language attd Linguistic 7?leory, I I. 
I,evin, B. ( \[ 993). F, nglish Verb Classes arid Alter, ations ~ A 
I'reliminary Investigation '" The University of Chicago 
Press. 
Marlin, S.I';. (1975). A ReJelence Grammar of Japanese, Yale 
University Press. 
Makino, T. (1991). Natural I,anguage Processing (in Japanese), 
pi).84, Oomu-sha lad., ISBN4-274-07658-X 
l:ukui, N., S. Miyagawa, aml C.. Tenny. (1985). Verb Classes in 
l';nglish and Jap,'mese: A Case Study in the ltlteraction of 
Syntax, Morphology and Semantics. Lexicon l"rojcet 
Working F'apers ++13, MIT. 
Pustejovsky, J. ( 1991 a). The Gerterative Lexicon. Computational 
Linguistics, 17.4. 
Pustgiovsky, .I.. (1991b). The Syntax of Event Structure. 
Cognition. 41.3. 
Sanseedoo. (1990) New Crown Japanese-l{nglish Dictionmy I?ev.5, 
Sanseedoo l+td. Japan 
Talmy, l,. (1978). Figure and Ground inComplcx Sentences, in 
,losel)h (hecnhcrg (ed.) Uttiversalx oJ' l htman Lan2uage : 
Syntax, Vol 4. Stardord. 
249 
