Semantics of Complex Sentences in Japanese 
Iliroshi Nal(ag~w~ Shin-ichiro Nishiza.wa. 
l)ept, of l¢,lecl;romcs a.n{l (,olnput(l " ' ' " I \]lguwe,' ,tg, Yokohmna. Na.ti{}\]la, I 15,\]iv(;,'sil;y 
('-ma.il: n~d<agawa., ,q}lill }{~naldab.dllj.ynu.a.(:.jl} 
Abstract 
The import;rod; lta,rt of s(;mantics of Coluplcx men 
tenet is c{q)Lurcd as relations ¢Ll\[l(}llg S(}III~LIII;JC l'ol{}s 
in subor{linat(; attd main (:lausc reSltcctively, llow- 
ever if thor{; can b(; relations between every l}a.i\]' c}f 
semmlt;ic roles, th0 ;I,lllOllllt, Of (:onlltutation I.o idet> 
tit'y the relatiions l;}iai, hohl hi the given senl;eiic{~ 
is oxta!cmely lal'gc, hi this I)a,l)er, for scnianl;i{-.q of 
Japa, ncsc c.olnpicx S0Ali,Oli(;Q We inl, l'o{hice liOW l)l'l:tg~.- 
n\];tl,i{; roles <:all,,d ()b,7(JrlJ~r ~ui(I molivaled rcspc{> 
tiv(;ly 1>o bridge s(;\[lt{i,\[l{;ic \]'(}l(:ts of subordinal.c and 
those of main cimises. I~y these new roles COliSi;r<-%ill{,s 
on the relations among s(;iilal\]{,ic/l)ra,g?~n~d;i(: rol{;s arc 
knowrl I,o bc aJlYlosl, local within subordinai,(~ or lily,ill 
{;l~iuso. \]\]l {}tiler words, as for I;\]ic SOl~l;tiiti(;s o\[ l,h0 
whol(; coinph;x S{;liLOIl(;e, the only role we Ml{}ul{1 ,le~d 
with in a, molivaled. 
1 Introduction 
Our a,ini is to f{}rin~tl\]ze {:(}nsl;Pailll,s l, hal. arc tl{;{;de(I 
t{} dcvcl(}p a. l)a, rs(;r I}as{'(I eli unilical, i(}ll ~ralllltl;ii: 
(called "U(\] >, h(,ql(;(~f()rth) so l, ha/, {}lit' l)ars0r (:ali {leal 
wil>h v;u'icl;y of types o1! S(Slllo{':Iic{ss ill :lal)an,'se. llow 
ever just parsing syntactic;dly is ,loL Cilough E}i' lial, il- 
rM la, llgll~tg(:: liit(|ersi,all(li,,g, ()11o ilill)(}rLalll, and ll(IC- 
cssm:y l;ask t;o 1){; doll(t} WIICII ~t l)a3'sor I)I:OCOSS{;S 3. 
{lis(:oursc in J{H)anos% ix {,It{) s{} called zero ana.i)hora 
resohltion. All of syntactic, sel\[ianl.ic, and pi:ag- 
rnatic constraints are i;o I)e h,volve(I to resolve z(;ro 
a, llal)llOlN~t, Of C()lll'S(?, mOlLie ()\[' Olllit,~(?(I ltl'{}lIOllilS al'(! 
synt, act, ically resolved. For instance, VP wil,h suffix 
~1~ iS llOt regarded as a (;l~nlst' hilt it {'.(}lijliliCt VI }. 
Th(;rcfore l;hc subj('.cl, o1' 1.h{! VI i wil;ti lc, whicli is 
ltossibly {}triii.ted \['rOlll Slll:fa, c{?, shoul(l (;or{d'cr with 
l.h{! sulijecl; of t.he SOiltOii(:(;. ()lie exa.inl}le is 
(1) Ilanako -w;J~ ¢1,~j sa.intlkilq;C -Tel'It7 fl;cl cold 
q52subj ina.{lo o siinc {dL 
window - ACC closed, 
ql~mako felt cold and closed tim window.' 
where both of zcuo sul)j{;cts ¢lsubj and ¢~,U 1 
refer t,o the scnt.cnLiM topic Iiauako u h/ tilts c×. 
an@c, one o\[ tilt; possihlc a.ccounts for 1.his interprc- 
t,,%ion is the following. Zero subject oF -Is phrase is 
\[ -I mini}boric, t l}rouominal \] or PI{O in GB tcrm 
\[Sells 85\]. As l.he result, ~bl~,bj is contr,)llcd hy the 
suhj{'ct d)2.,,~,; of th{; ma.ill VP, which is ;\[Is() zero sub 
jcct. ~%,@ is, iu Clll term, \[ anaphoric,-I- ltronom - 
inal \] or pro. q'h{' senl.cntM topic Ilanako is the only 
ll()ssihlc a.nl.eccdc,,t of this zero suhject i, this exam 
l,lC. Ilowcver, in conq)lex s(;ntenccs, I.liil,gs are {luitc 
dilli;rclll.. Consider l;h{; lb\[h,willg s(ml.clice. 
(2) il~ma.k{> - wa \[ {l'l..,~,; sa,rttt 
-'l'Ol>l(J \[ I'eeliilg cold 
-g~l.-t.a node; 
1}chavc(l like I)o(:a I/S(;\] 
<IJ'28.bj lnado - o silii(>l;c, yal, 1,a. 
whtdow- A(7(7 close gave. 
I. <Since lla.nal~.o I}cll~wed like feeling (:old, I 
closed the window.' 
2. <Since \[ Itch;wed like li;cling cold, Iianako 
clos(;d tim window.' 
If contextua.lly wc (:an I.ake only ltalmko and tit(; 
Sl)(;aker of this senteHce as candidates of anl.ccedent 
of (/)l<subj eli" {i)'2,~ubj, inl.uit.ively the following two in- 
I.crl)re(.ati(}ns ;it'(; equally likely. 
a. ¢ i,,o,i - I lan;Lk(), q52.~ubj : : Sl)Cakcr 
h. dJt,,~,; speak0r, (/JS~ubj :- II;uiako 
'l'hercfore (llisub j &lid {//Ssub j ~/,l'C hotli \[)re. |11 fa(;l. 
this I'act is well klloWil lHlll)llg~ ,I zq)a.ncso \]ingllisl.s, i.e. 
\[Sells 85, 'Pdcul)o 87\]. As ;t roslillg zero iHtaq)hora, 
rcsolut.ion o1' con,I)lcx s(ml;cn(:c is l,ot only to I)(; 
don(; sylll;acl;ically, t)ul. also 1;o })e done l)ra.g;lnal;ically 
a.n(l/or SCil\]a.ni.ica.lly. Oil{,. o\[' th{. i)roluising can(ti(la.tc 
ror this is the (;cnl.cl'hlg tdieory \[Ih'onnan eL a.l 87~ 
Walker 90\], To apply the ceill;ering l;lieory thai, is 
orighially for :~ SC{lilCU(:e of s{mt;eil(:es, liarllely dis- 
co'~H",';c, WC i'egm'd Lhc sul)oi;diuMx; (;lmlsc a,n(I the 
Ill,till (;lmisc as ~t scgflrlOllt or (liscourse rcsl)o(;tivoly. 
Moreover llaimko who is marked 1)y <wa' is regarded 
as l;he topic for t.hcse two cla.uses. Then, the l.oplc 
'lhmc(fforth, +\],$$$ lile~lilS; 7,1!1"O $$$,., where $$$.. is ell, her 
grinmna.l.h;al, so.,ilittil~iC {ir l)\['&ginatlc role. FoP hlsL;ti,cc, (/}.~ubj 
I\[I('.D~IIP; Z(~l'O Sll\])j(!cl~,(/Jagl lllt!&litl %(\]I'o fsq(~7t, T, f./Je;L'/) llll)illlS 7,01'O 
L3JIJtsTit"lI.C¢'7" , ;LIII'I SO l}}l'l}l. 
2 qlanako' is ;i typical giN's li&iil(L 
679 
Hanako is the strongest candidate for the backward 
center of the subordinate clause. Therefore the hack- 
ward center of the subordinatc clause is tlanak% arm 
consequently zero subject qh,,,bj refers to Itanako. 
By the same way as the subordinate clause case 
is dealt with, the zero subject o\[' the main clause 
4~z,~bj is known to refer to tlanako, too. This re- 
sult is neither interpretation a nor b shown above. 
Another candidate is the property sharing thoery 
\[Kameyama 88\]. In her theory, since the both of 
zero subjects share the snbjecthood, both of them 
finally are known to refer to Hanako that is the topic 
for both of these clauses. Therefore the prol)erty 
sharing theory also fails to account for the iutuitive 
interpretations. 
Then we shift our attention to more microscopic 
one, ill which ,roughly speaking, the important part 
of semantics of complex sentence is tbrmalized as re- 
lations among semantic roles that appear ill tile main 
clause or the subordinate clause. At the first glance, 
the constraints about these relations are not local in 
terms of mMn or subordinate clauses, hr other words, 
semantic roles that appear in subordinate clause and 
semantic roles that appear ill the main clause seem 
to be directly constrained by the constraints of com- 
plex sentence. However, looking more carefiflly, we 
find that the constraints of subordinate clause and 
tile cons(fronts of main cla.use are represented as lo- 
cal constraints by introduciug the new notion of mo- 
tivated which is characterized as a llerson who has 
euollgh reason to act as the lllain clause describes. 
More precisely, moZivated is one of tire pragmatic 
roles that appear in a subordinate clause, and the 
constraints in subordinate chmse are stated as iden- 
tity relations between molivaled and other seman- 
tic/pragmatic roles appearing ill subordinate clause. 
Therefore these constraints are local in subordinate 
clause. The constraints ill main clause are stated 
as identity relations hetween molivaled which con|es 
from subordinate clause, and other semantic roles al / 
pearing in main clause. Therefore in understanding 
the mail( clause we (h)u't have to be care M)out se- 
mantic/pragmatic roles in subordinate cla.use other 
than a molivaled. In this sense, the constraints ill 
the main clause can be treated as almost local con- 
straints of the main clause. 
The next question is how to represent tile seman- 
tics of complex sentence in feature structure( called 
lPS henceforth ). l?or this, we shouhl write down 
the constraints about these relations among seman- 
tic/pragmatic roles ill a feature structure formalism. 
Due to the space limitation, in this l)aper we ma.inly 
pursue the constraints about semantic feature struc- 
tares. 
2 Hierarchical Structure of 
Complex Sentence 
We pay our attention to the general structure of 
3apanese utterance which is helphfl to rel)resent 
semantics of complex sentence. Several Japanese 
linguists have Mready proposed the general struc- 
tm'e of Japanese utterances \[Mikami 53~ Minami 74, 
Takuho 87, (~Ullji 89\]. Mikami categorized clauses 
into three (:lasses, namely 'open', 'semi-open' and 
'closed.' '\['his categorization indicates how freely the 
content of clause interacts with the outside o/' clause. 
For instance, they arc categorized by the degree of 
possibilities of coreference between zero pronouns in- 
side the subordinate clause and nominM or topic that 
appear in tile main clause. Following Mikami's idea., 
Minami proposed four levels, namely level A, B, C 
and D which correspond roughly to VP, proposition, 
sentence without eommuni(-~tion mood and utter- 
ance which takes into a.ccount a hearer, respectively. 
\[Takubo 87\] divided level A into two levels. One of 
them corresponds to VI', the other corresponds to 
VP + a certain kind of subject which is called "ob- 
jective subject." g~unji proposed the more detailed 
structure, in which starting from predicate, say, verb 
and adjective, objects, voice, subject, aspect, tense, 
modality, topic and mood are or might be sequin> 
tially added to make all inIbrmatioually more ful- 
tilled sentence component, l"iually, it; ends up with 
all utterallce. \[n GIlIlji's strllctlll:e, solne node ('all 
have more than two (laughter nodes to make more 
coutl)lexsentence. Following them, tim structure of 
the so called (chlase level) complex sentence is the 
following shown in Fig. I. 
(ltl,eraace 
,J udgerlleIlt Mood 
Topic CJounnent 
Event Modal 
Suh-Clause I)roposition 
Conullel/t Conjunct Process 'Fense 
Action/State Aspect 
Suh,iert VP 
Object V l' 
l'redicate Voice 
Figure.l: The hie.rarchical structure of ,/apauese ut- 
terances 
In Fig.l , Sub-Clause and Conjunct mean subof 
dinate clause and conjunctive partich: respectiw,Ay. 
Note that Fig.l represeuts not only the hierarchi- 
cal structure but Mso the word order of a corn+ 
plex sentence in Japanese. The structure is al- 
most tile same as Gunji's structure except for explic- 
itly showing complex prolYosition , subordinate-clause 
680 
and conjunetiv<>l}artic\]e tha.t are newly added t(} deal 
with COml)lex se\[Itenees. Note that '(;ommenl;' a.p- 
l}earing in 'Sub Clause' has the same structure ~ts 
'(',olnlilCllt; } al)l>Cal:ing just below 'JudgelHe.ut'. r\['h~tl, 
is to say, '(~omment' is re<'ursivcly delh~ed. Ih}wever, 
ill 1)raetic<'=, the m(n'e Lhe level <}\[' del/I;h (}1' rec,lrsivcly 
apllearing '(~Olnlrlent' is, the less (;(}lnl)rehetlslble the 
sentence, is. 
3 Subordinate Clause 
In this section, a.t first we show tim predicate cat.e- 
p;ories used in the subordina.te clauses that we deal 
with ill this 1)al)er, in Tal)le.I. lu each category of 
2,;I,4,5 a.nd 6, exists there+ a i)ersoll who is alfecte{l 
by the situation descril>e(I I)y the sul>or(li,ml.e clause. 
On the eontra.ry, in earl'gory 1, there is not neces- 
sarily an exltliciL aIfe{:ted l)erstm. In our tll{!ory, this 
affec.l.ed iterson plays a, key role for selnantics of eom- 
Iflex sentence. As tim re, suit, in general we cannot 
derive a useful result for ca.l,eg(n'y 1 in our theory. 
There17orc we d{}n't dea.1 wiLh category 1 in this pa- 
l>e r. 
At this momeut, we sh{ml<l expl;fiu the tml.urc 
of the. so called sul).iect.ive l>redicate Inentione.d in 
Td)le.I. Ill short, a. sul/.iective pre.{li(:ate descril}e.s 
tim e.%'perie'ncer's inner state which can exclusively 
\]>e kn{iwn liy I, he ca:pcricvcer him/herself. 
Next wc \[5eus on verbal suilix :larU. Firstly we 
show .<larU's syntax. (:aru is the= prcseut \['orul aud 
its reel. form is gar. Therefore inlle.eti(ms are as fol 
hiws: #(t'c:vc}#aP-i, et(;. In addition, .q~u'u has &ll 8.1- 
lophonic root I'(}rnl gal an(I, .qag=la(ltast-forln), gal- 
leiru(progressive..fortn) a.nd so on are derived frour 
gal. Some. (}f these forms will a\[}i)e~u' in our ex- 
atnliles. Next we talk about th{; sere;reties {if garu. 
Uaru roughly means "show a. sign of" or "ltehave like 
7F+ Als,, i,, ss\] its se,,,..ti(-s is 
informa.lly explained, however our \]>r{/p{>sal is 1.o for- 
malize tar'it's semantics in U(.I (>r more generally in 
computa.tiona.I linguisties, l"or I,his, tirst of all, we 
introduce a. new pra.gmalie role called observer. 
Delinit;ion 1 (OI}server)Observer is :l person 
who direclly observes or is iudireclly in.formed lhe sil- 
.uahon described by lhe p'roposiliou parl. 7'hcrefore 
ttn observer has a ccrlaiT~ evident( to I,(: cottvinced 
lhal lhal sihtalion acl'ually haplicns. 
-~-~ lum-sul)jcetive in'e~dieat(~ 
2 I subjective verb 
3 \[ sul)jeetive adjective with<>lH, verbal suffix garu. 
4 I subjecl.ive a(Ijeetive with verbal suffix gar'u. 
-~--\] verb-t t~t-g(trlt 
(l>ehave= as s/he wants to rl' ) ____._+ 
(i I transitive passive anti iutransitiw~ i>assiw~ 
(a<lversity i>assive.). 
Tal)le 1: Ih'e(licate C, ategoPics 
Although 1,his notion of observer'shares a large part 
with PIVOT of \[iida-Sells 88\], our notion of' observer 
is introduced only by garu. Therefore it is much nar- 
r<}wer notion. As y{,u will see later, this ne.wly in- 
l,roduecd role is playing a key role which bridges se- 
mantic roles of subordinal:e. <:lause t,o sclnantic roles 
of main clause, 
#ks for an observer introduced by garu, one o17 the 
widely known consequcnc.e aboul; the nature of suD 
jective predicate is the following, ht a sentence, if a 
subjectiw" adjective is used without being fo\[lowed by 
a verbal sulfix gar'u, the ez'periencer of the subjective 
adjective, should he tit<; sl>eaker of the sentence. 
The next thit,g we should do about a newly intro- 
duced notion of observer is to make (:lear the way to 
deal with it in FS. First of all, in our t"S, a semantic 
cmltent:Sl,;M is I)asically a sea. (state of alfa.ir) form 
of sit.uation semantics. Ilowever we use semantic role 
like "agent", "patient", "experiencer", and so on, as 
argument roles of sea. Since= an observer observes 
the situation which is characterized by a sea, if we 
know that there exisl.s an observer, the observed soa 
is eml)edded in (/bserving situation, which, in turn, 
is elnl)edde.d ill the whole semantic (;otltel|t. In this 
SeAlS(+., the observed sea's itr~ttttlellt role is observed. 
I~llt ;iS \[~,1" ;IS W(? hl/,ve IIO (;()llfltsiOtl, we omit role llatlle 
'ol)served' henceforth. A tyl>ical scltcma of SI",M <if 
I"S of this l,yl}e is {,lie following. Note that we use 
:Iar*t as a value o\[' the rela.l, itm feature meaM I>y 'rel.' 
The Et@ish gross of this relation garu is %bserve.' 
rel: garu 
ol)servel': ~\] 
rcl:lL 
(:l) SI,;M = ~gent: \[k~ 
so~: e~perie,,{,cr: ~\] 
,t <e,tt: 2 l 
Now we explain l.he semanti<:s of chmse wlfielt c<m- 
sists of sul>jcctive adjective with 9aru or la-garu., 
l.\[laA, ;+ire ill categories 4 and 5. These categories' 
\[\>rms are "(/:<~p P-garu" or its past form "4e:cv \]'- 
gat-/,a.", where 1' is a subjective adjective (category 
4 iH '\['ahle.l) or is a verb followed by la-(lar (ca.re- 
gory ,5 in Table.l), and </)<,r is the c:rperiencee of P 
which is l)ossibly zero. lu these categories, there ex- 
ist. el)servers who are not the+ e:cperiencer of 1', and 
observe, that exl>erlence. The gEM feature of "qSex P 
I'-garu/gat:ta" is the followiltg. 
ol,.e, ver:D wl..'e \[\] ¢ \[\] 
(4) ,'el:P \] 
where " -/ " means "not token identical." 
lit our l"~q, eoustra.ilfl;s 17or toke=ns like\[~ are written 
with "where" as shown in I.his FS. Since constraint 
satisfaction method in UG ha.s been and is devel- 
Ol)ed hy lnalhy resea.r(;hers rece=ntly i.e. \[Tsuda 91\], 
681 
our theory will be able to be iml}lemented in systems 
like theirs. 
If the sentence finishes just after "garu/gat-ta", 
the iml)ortant points arc l) an introduced observer 
is the speaker, and consequently 2) the es'perienccr 
cannot be the speaker. If a clause with "garu/ga.t- 
ta"is a subordinate claus(;, the ca'per, cheer cannot be 
identified with a semantic role corresponding to the. 
subject of main clause or higher clause. 
As for category 2, subjective verbs like "ku- 
rusimu" (feel sick) and "kanasimu" (feel sadness) that 
describe subjective and/or emotional experience in 
verb form, are used. Like the case of gar'u, an 
observer who observers the experience can be ill- 
troduced. However this observer is not obligatory. 
Therefore unlike the "garu/gat--ta" case, the expe- 
riencer also can be an obligatory senrantic role of 
higher clause as well as the speaker. 
4 Complex Sentence 
4.1 Feature Structure 
According to the hierachieal structure of Japanese 
sentence shown it\] Fig.1 , the essential l)art of hierar- 
chical structure of the following sentence (5) is sllown 
in Fig.2 . In this figure, the structure just below e~ch 
proposition is replaced with the corresponding parts 
ol" sen,elite, 
(8) \[~:,p smnu -gat-ta \[ feel cold l)ehaved like 
node, ¢,nll mada-o sime-ta. 
because \] window-ACC closed. 
'Since 0~,p behaved like feeling col(I, d%tt closed 
the window.' 
Coml)lex l}r(}l)osition 
Sul)-Clause Proposition 
Cfiotl'ulleut (~otljuIlct mado o sime-la 
I I 
I }roposotion node 
I 
samu-ga t-ta 
Figure.2 : Iliexarchical stru{:ture of (5) 
Basically the eml)e(Idi\],g str,tcture of FS {:(),'re 
sponds to tile hiexarchy shown in the hierarchical 
structure Fig. l . To grasp the image of the relation 
between a, hierarchical strH(;ture and the corl:esl)oltd 
ing I!'S, we show an example, of FS of the above com- 
plex sentence (5) analyzed based on this tfierarchical 
structure in the following. This 1eS is the result of 
tile unifi(-ation between the FSs of subordinate clause 
and main clause, where the contents of syntactic fea~ 
t.ure IIEAD , namely ~ is omitted. 
MOR, PII: 'satIHl-gat-ta no(le,mado o sime-ta' 
IIEAI}: \[~\] 
sl~:rv,: \] matrix-sere ~A~ A \[\] :fi \[\] 
,:el: node 
motiwmxl: \[\] 
l rel: simc agent: \[\] object:window tense:past 
rel:garu l observer fc!\] 
soa.: \[ rel: samu-i 1 soa: | experiencer: \[\] 
L tense:past 
where English grosses of relation name is the 
following: sime:'close', node:'because', samu-i:'feel 
cold'. 
The key point of the semantics of complex sentence 
is the role lnotivated that appe.ars in 
which corresponds to the content of the subordinate 
Cause. '\['he role molivalcd is the link between the 
content of subordinate clause and the main clause. 
Semantically motivated is characterized as the fol- 
lowing. 
Definition 2 (Motiwlted) Motivated is a person 
who is aj\]~cled by the situation described by the sub- 
ordinale clause deeply enough Io feel or acl as the 
main clause describes. 
The important and indispensable, part of seman- 
tics of c{nnplex se.ntence is, roughly speaking, the 
relation between a subordinate clause aud the lib'.till 
clause. But if you look more closely, I;his relation 
is actually tim relations among semantic/pragmatic 
roles appearing ill the s,fllordimlt.e clause aM those 
appeariug in the main clause. The newly in~ro(tuced 
role of motivated gives {.he most important chic for 
this relation. Therefore, in the rest of this paper, our 
effort will be concentrated il,to whom a motivated 
refers {,o. More precisely, in FS, our main concerns 
are which sema,tic role in the SIqM of subordinate 
clause the molivaled can or cannot be. unitied with, 
and which semantic role in the SEM of main clause 
the molivalcd can or cannot be unified with. 
4.2 Constraints 
In this subsection, we propose the co.straints on 
complex sentence. For t;his, al; tirst we categorize 
the relations between subordinate, clause and main 
clause t)ased on their semantics. They are divided 
up to many types of complex sentence. We show the 
most important and tyl)ical types in 'l'al)le.2, where 
SC an(l MC mean 'subordinate clause' and 'tnain 
clause' respectively. In this gable, the first; column 
is for a name of sentence type, the second column 
indicates a rough meaning of the relation 1)etween 
682 
...... , coitlplex SO, tlt(~ll(:(~ 
S ~ t;+IAIS()S MC 
"= ,' Althongh SC, M(7 
3 I If SC (,hen MC 
4 i ~,¥ hcn/~fft(;r/1) e fort 
(;L<: SC, MC 
nod(:, kar<l 
no'hi, 9a, kcrcdomo, 
/c'm o, i-It, 
i- l It lit, i.. ~l a g a ~'a 
to, nara~ tara, rcba 
toki, ato, 
Ifl(tC~ ct,(', 
Tat)l<, 2: (\]\]aJlse Adjuncts 
~ubordiual, e, clause S(\] a, nd lu;til/ clause M(\] <)\[' COtll- 
plex SOiltei~co,> ;.ul(l l./lc thir(I c,o\[Itititi ~h('lws .\]apallel-;e 
colijunctive parii(:l('s use,(I 1,o r(;i)rescul, ;l, l, yl)e of COlll- 
l>l<;x S(HII,oIi(X) hi th(; ~;l~iii<', row. 
q'lll'eC VP ad.iilnct, s , It, lltlu, and ll(l(jfll'fl, 3£(; iiSil 
ally used to express o,v(Hi{;s o(;Itrritig .'-;hlltlll,~tllt':OllS\]y. 
l\[owcver, il" they are used wii,h asl:iC(:tua\[ su\[lix 7 
whi<:h lll0gttls lierf<:ctive, for iliSl,au(;e 7-1t(tfjttFa> lltcy 
axe llCg~trdcd a,s cla, ilS( ~. COli.iilllCt8 ;till\[ al;0, to t)e illl,er. 
prcted a,s 'although'\[Mina, nii 7d\]. We d(>ll'l, (loal with 
tyl>c 4, t)(;Ca, ll~() ~t l,elnl>oral a(lverhial clause illS( (\[e.- 
scrlbc, s aAu (;VOlll, (,\[ial; OCCtII!V~ h()\['Ol'Q, shil/tll,3~m~oitsly 
(:ill after aJlol;he, r eveni, which in (te,~crilt<~d Ity tile nlaii~ 
(',\]ause. Thercfc)re genera,lly we don>l; (~xl:iecl, cssentia,\[ 
hiforin.%tioil for rei~d, ioli~ itlliOlt/~ s(;illali(.ic t'olc~ a\]) 
p0,arin~ in a,dv0,rl)ial or iila\[li clause ft'()ill lihi6 I,ypc of 
so Ill.ell co. 
Now we fo(;tis oti type 1,2 all(t ~l> where ~t m0= 
lilJ(t~cd plays I;he key r()le in the ('()li.~trahiCs. in 
Table.3 we sliow the (',OliSl,r{titt(;s t\[i3,(, say which Se 
uiantic/pragina,tic role of sllbor(\[inal:e claus( can h<; 
a. moil'rated. Tahle.4 ,shows which sen(ant, it roic ol; 
ma.iti clause <:a.n l:ic uni(i<xl with the al, otivc~ted, hi 
these i;a,l>les, the first (:ohlinn of the lirst row in for 
oollsI;r;tinl, IlalllCS~ the scC, olld r, ohllrill ~h(:iw~ a set (:ir 
~eriL(;i/ce typos for which the colistra, ilil,s s\[iOWtl hi 
the+ secoli(\[ row apply. The i,hir(I cOilltiili of Tab\[<'.3 
,<-;how~ predical,e l)a,lteA'ils (if' sui:ior(lhi~d,e clatlS(!, and 
i,h(~ thh'(\[ cohlmn o\]' Tahle.d ~hows ~etiialit, ic c;l£o. 
gories of prc'dicatc of ltiahi claus(. I~l)l ' I,\[leni, coil 
8Lra+ilil;s Wi'it;t(~li ht t,\[i(; sec()n(l row apply. Note /,\[iat 
all ()1' l,hesc coiislA'a, itlts hi Tal)l(,.'.l al'() lo(al in a ,~llbor 
(\[ina(,c ('la, tlSe~ I)cca.usc, both ,~idcs o1':: <>f c(msl, rahll,~ 
are roles (l\[' mil)ordhial,e c|a, use. hi (;as( of ~ul:ijc, c- 
1,iv0 a, dj<,'<:tive wil, houl; garu, l, he, COll~traiill, <lnolitlale.d 
-- erqmdencer ' hohls ~dso for type I ex<:cpt for tit(; 
case whe.re directionally auxiliary w'~l'D "yaru(.~ive)", 
"kurcrla(I)c given)" are used. Analysis for l.hese ca~e.~ 
in Oil0 O\[" ottr flit(Ire t)roi:iJenl. 
As for '\['abh;.4, stat, e + is a slat,c, e×c(!itl, |7)r the/:a~c 
that thct'c exists a third l:iarty who ix a molival((I 
l>UtS t,hc (>/EpCl'iClI, Cff*f" iliLO l,ha, i, s(;ai, e. For ins\[,attce, 
tim c,;e\],.el"Tc'ntel' is periuii, t;ed 1,o <1o s()iuol.\]lill<l~ I:iy Ihe 
ln.ol'ival(:d. ~hu;e hi t;his kind <>f case l,hiltgS ;n'(~ quite 
conipli(',at, cd~ wc olrtil, il. |lel>e t)ccaAls(~ o\[' the, lhllitC(I 
spaoc. (Jol~sLr~hil,8 hi Tabh;.4 are also local in ~c tnaJu 
el;ms( I)ocaus(; (very se, ina,ill.i(: role thai; at)peare~ hi 
the ri<~liiJia, n(I sill( or tiic c(:iusl;ra.hil, s i~ delhi(;d within 
\[_' Im m<! 7 - - - 
\[ __ ('.oil si,r ailll; 
......... veri) I ~ la 4 gala 
moli'valed :: observ(>/r 
~-7~ ~,,1,.i,;c(,iw. ,,4i<~ct.iv(~-i <~,,.,, 
L-- ~_ verb -t- la "\[- gartt 
l,~,~ ~v tT~., l~ 
Ittoli'ltale(l :: obscrver V c.x\]2ericllf;(>.r 
~--- \[ sld, i;;ctiv(; a(lj<;cl;iv<~ (without .qt,,', 0 
molivalcd := cxperie.necr 
~vc (without garu) 
molioalcd ::: e, xperic?tccr V obscrvc'r 
"motivah>.d :: (dlT~(:h~(I 
'motiva&d :: a lf(ct(:d :if a.lJ'e( h:d cxisllS, 
:: p(tlicnl : otherwise 
wher0 i, \[1 ttlO, lLtlS iI, \]l~tltle til' ("~Lclt COltSl, l'~'tilli;. 
Tal:ile 3: (JouMa'ahll,s ht ,lqul)ol'(lhia, tc (\]\[a, ilSC 
,ta,tTe~_ i,ype_\[ __iir(71{~il'(;_ ,, {;aTe~o,'y : : ,,0 _ - -- 
coitM, rz~hll,~ \] 
:, - \[ "- _ _ 
P s li\]:i (:lau.~;c:~l';~(l:,/to!/'_va!f d :: agc,,I \] 
_stil>;{iiTu~s(~:Si<',~4:~t,;;iiv(,t(:TI ::= e:,:l>~7.i(:~,{7,~./ j 
Table 4: (<.(msl.raints in Main Clanse 
(.lit (ttaiti cl3.1t.<J(;. Nee(lless i.o say, the hdhicnce \['roni 
~L sul)orditiat,(~ (:lailsc COltlOS ()lily via i:(Jlc m.oliv(ttccl, 
In l;tle, reM, (:if this s<~cl;iou we, Mlow the cxanil)les 
l;liat cxenil:ilify these (:Clll~t;ra,ill(,s. it 
I+h'~t~ wc tak(, (<5) of type I. The c(mstrahll:s I.o be 
applied are SI and M 1 as you kuow front the contellts 
of sub(ni(lhta,te and ill(till cla.usc. Ily (:Oli/bili~,ll, ioli o1' 
\[ ;tnt\[ M l, zero a.Cl(!ltl o\[' llla{ll (:\[3MS(}:C/Ja!/l in l, hc ob- 
server o1' the .<di, ual;ion des(:rit>cd I)y l, he 8ut)or(iinai:e 
clau~e, wiier() (lJc;(,lj bchave(\] like fe(~lhlg cold. This 
inl,erln'el,al, i(:in coincides wifll n;d,iv<;'s ititiii(,i()ii, 
I,()ok ;tt, I,\]lC I'ollowiug pair o\]" c×;Itlil)lc, 
((i) \[0<':,:r I¢lu'usi -.~al,-I.a il<)tli\] \[ Ib, el hat\[ I)eh avo, d \])tit\] 
kekkyoku q$,.~st 
al, last 
kttSltl'i - O IIOtil ~l, llll,k~t,t-t,%. 
lUe(iic, hte- A(;J(\] drhik ItoI,-PA~T, 
'Althoul,;h q~:,~:r I>eha.vc, d like, feelin~ ha=d, (/~-sst 
didu'l take a ni(,diclue at last.' 
\[¢'::w llokori t,a -<gat-t a 
(7) \[ stay want b(;haved like 
uotg\] kekkyol(u q%~t oi dasi-ta 
but \] tiredly ti)rcc.d out. 
3'\]'he ex;tnli)les  ho',vui I)elow at'(! ;l Lip ¢)f i(x!lmt'g ',Ve ;I.C| H 
ally ;malyze(I, of com'se. We gaihcr the (latz~ al)out \[ud, lvc'n 
in(.uitiw: interpret;tt, ion fl'OllI lllOl'(! I,h;~D twenty mtt, ivcn iH'()I ll( 
;IAlI hln's, 
68.3 
'Although ~b~.~p wanted to stay, qS<qt finally 
forced him out.' 
In both of (6) and (7), the m.otivateds of subor- 
dinate clause are constrained by $2, namely moti- 
vateds can be either q5~, v or the observe," of subordi- 
nate clause. Constraint M1 says that in both cases, 
qS~.qt is unified with the reel\[rated. Intuitively in (6), 
qS~gt is ~b~,.p. On the other hand in (7), G qt is the 
observer. Both of these interpretations comply with 
constraints $2, and M1. 
\[q51exp atui node \] (8) \[ behot because\] 
q52exp komaru. 
be in trouble. 
'Since it is hot, I am in trouble.' 
Intuitively q$1~.,.p corefer with q52<.:~ v. This inter- 
pretation is expected by eonstrMnt $3 and M2 that 
apply in this case. As you know from these exainples, 
our constrMnts are not strong enough to identify the 
antecedent of qSan, uniquely, lint makes safe inter- 
pretations. Moreover disambiguation done by these 
constraints is useflfl for further inference that will be 
done with comnionsense knowledge or with a special 
vocabulary like 'kekkyokn(finally)' used in (7). 
In case of $5, namely iritransitive passive or ad- 
versity passive, it is well known, i.e. \[Gnnji 87\] that 
tbere exists a person who is affected by tile situa- 
tion described by tile passive sentence. An example 
sentence is the following. 
\[~Jaffect trirria- ni sin -are 
(9) \[ wife be dead -PASSIVE 
-ta noni\] 
-PAST but\] 
~exp kanasimi - me -st nat. 
show sadness not, 
'Although his wife had gone, 05~:,,~> doesn't show 
a bit of sadness. 
The semantic role of this a.ffected l)erson , ill (9) 
zero role:Gqj~t whose wiD was dead, is ~tl'i ajfected. 
Tile intuitive interl)retation that 6~*'i, = Gd.r~t( = 
motivated), is expected by our constraints: $5 of 
Table.3 and M1 of Table.'\]. On the contrary, lit case 
of $6, namely transitive passive, gellera.lly we don't 
have an affecled, ltowever in seine context, a transi- 
tive passive form lnay require the role affeclcd which 
is inherent to adversity passive. For instance, 
(10) ~a.Uect saihn- ga. iillSUni 
wallet- SUFI3 steal 
- are - t a 
-PASSIVE -PAST 
'd)a.ffect's wallet was sl.olen.' 
hi this case, a. person whose wallet was stolen is 
not explicit but regarded as an affected. Another 
case having an affected is that a rela.tional noun is 
the subject of trlmsitive passive. Then a person who 
is in the relation exl)ressed by the relational aolln is 
thought to be affected by that situation ,too. Here 
we take 'lnother', 'father', 'daughter', 'son', 'super- 
visor', and so forth as a relationM noun. A couple of 
exanlple sentences are tbe following. 
(12) 
\[ kobnn -ga yar -are 
(ll) \[ henchman -SUB3 attack -PASSIVE 
-ta node \] 
-PAST because\] 
~b<,nt sikaesi- ni it-ta. 
retaliate go- I)AST 
'Since his henchman was a.ttacked, the boss re- 
tMiated.' 
\[ kobun -ga yar -are 
\[ himchman -SUBJ attack -PA.SSIVE 
-ta noni\] 4b<q~ te-o komaneite-iru. 
-PAST but \] did nothing. 
'Although his henchman was attacked, the boss 
didn't retaliate.' 
qJ,<.qt who retaliated (1 1) (or didn't retaliate (12)) 
has a certain relation between the henchman who had 
been attacked. For instance, q$<</t may be the boss 
of that henchnian. In ill), since constraint $6 of 
TaMe.3 and M1 of Table.4 apply, ?5ant is an affectedof 
attacking event described in tile suboMina.te clause. 
This interpretation coincides with native's intuition. 
In sum, with these constraints, a constraint sat- 
isfa.ction process in {JO based parsing can be done 
locally and consequently very efficiently. In other 
words, l)rimarily a constraint satisfaction process of 
a subordinate clause can be done within the analysis 
of subordinate clause, and tllat of the main clause can 
be done within it except for using motivaled whose 
value has already been constl:ained in the subordi- 
nate clause. 
5 Related Works and Conclu- 
sions 
One of the relevant researches to ours is JPSG that 
has l)een developed I)y Gunji\[(hirtji 87, Gunji 89\] 
and is further studied by tile ICOT working group. 
Ollr focus is a l/lore praginatics oriented one tha.n 
,IPSG is. Many Jatlanose \[inguisl.s have ah'eady 
done the enormous aliiount of I)asic observations 
aim proposed linguistie IJleories about the plie- 
liOlrlena, we deal with in this paper \[Mikaini 53, 
\[(till() 73, \](lllIO 78, Ohye 75, Minaret 74, "\]'akul)o 87, 
Tera.l\[iura 84, Terali'lllra 9(), Saito 92\]. Of ('.Ollrse Ollr 
research is based on their works and obserwrtions. Ill 
\[Ohye 75\], it is said that if garu is used ill a subordi- 
nate. cia,nse, the sill)jeer of I;h('. Inain clause is not the 
experiencer of the sut)ordina.te clause. In \[Saito 92\], 
she says tllat 1) a cognizcr that (;orresponds to our 
observer is introduced if .qaru is used, and 2) if an 
observer is introduced in the sul)ordinate clause, the 
mentally responsible lie*son appearing in tile main 
clause is identical with the observer. In linguistic 
l)}lenomena, these observations are similar to the 
684 
constraint we propose here. So what is new? The 
answer is I, hal,: 1) We explicitly stat, e the semanl.ics 
of complex sentence as the relations amot~g semttnt.ic 
roles. Namely, sittce we use scmantic/pragmal, ic roles 
instead of grammal.ical roles in constra.ints, our <-on- 
sl;traints (;311 ~-t(;('.Ollllt for geFo tLllO,l)hoI;iL ill a. s(:ntel.:e 
where tit(: main ela.use is l>assive where an agc)tl ol: 
an exl)er.iencer is not; necessarily l;he subject, like the 
following example. 
113) 
Taro -wa \[ gakkou e iku-no -we 
-qbpic \[ to school g<>NOM -ACC 
iya -gab ta. node \] 
hate behaved like because \] 
4,~11, @,,t okor :are -ta. 
scold -I'ASSIVE -PAST 
'Since Taro l>ehaved like hating to go t.o school, 
he was scolded.' 
where the intuitive, reading is the followiug: era', 
thai, is zero subject, refers t.o 'l'aro, and ¢,st, that is 
not the zero subject, refers to Ta.ro's parcnt.s who are 
the observer and molivalcdof the subordin;tte, clause. 
2) We formalize this t.heory in 111~ Ibl'malis,,t, even 
though the details are omitted due to the space lim- 
itation. 3) We find that l.hc constraixg.s of complex 
sentences arc, actually local ones. This localization 
el eonstra.lnt was found hy hltroducing new prag- 
m+d.ic roles observer atl<l ?~toliv(tlcd, and is extretr,ely 
iml>ortant tbr ellieiency of UG based parsing. '\['his 
localization also makes l.he prol>OSe.d constrainl.s be 
coml>osil;ional ones, I+ecause in l.he case el deeply <un 
I)eddcd complex scnteuce, l;o identify the rel+erenC of 
each m.otiva¢ed that, bridges I)etween a. subordinate 
clause and it;s IFlailt clause, the constraints we pro- 
posed are resolved with COmlmtation confined within 
each clause. 
Analysis of case in which a directional auxiliary 
vc, rb i.e. 'yaru','kurcru'is used is left as the figure 
l)roblem. 1,'imdly, we implel,ented ;t Japa.lmsc law 
guage ttn(iersl;anding sysl,etn l)ase(\[ ()it l,hc I, heory we 
sl;al,e ill +.his pape.r, but chic I,o t, he space lhnital, ion 
we will report t,he <letail of itnple~inenl.ation in other 
l>lace in the near \['uturc. 
References 
\[l\]retman ('% a} 87\] I~IX-~IIIHI, tl, S+, ~'l, W~t\[kcr Fried- 
man and C.P(>lhu:d 11987). A (3(;nl,ering, Ap 
preach to l'r<mouus. 251h Ammal Meeting 
of ACL, pp.155-162 
\[(.:u,,:ii 87\] Gunji, T.(1987). :1 a.l,a.,,ese Phrase St.r,,c- 
tm'c (\]('ammar. ll.eidel, I)ordrecht 
\[Gunji 89\] Guuji,T. (1!)89). lh~lev;mce of the For- 
realization of Phrase Struct.ure Grammar to 
Mechanical Language Processing. lh!l>Ort Hf 
'l%kute~i-Kenkyu, Ministry of I,~dueal.ion and 
Academy 
\[lida-Sclls 881 lida,M, aud P.Sells(1988). Discourse 
Factors iu l, he Binding of zibun, in Japanese 
Syntax led. W.l'oser) CSLI, Stanford 
\[K~..,,.~y+.+,.,ss\] I~.,.e>.,< M. (1988)..l,~>,,e~se 
Zero Pronominal Binding: Whc're. Syntax 
and l)iscourse Meet. in Japanese Syntax 
(e(l. W.I)oser) CSIA, SI;anlbrd 
\[Kat;agiri 91\] Ka.t.agiri,Y. (19911). l'erspeetivity and 
.Ial)anese Reflexive 'zibuu'. in CSM Lecture 
Notes No.26, Situation q'heory and its Ap- 
plications VoL2, J .Barwise el. al eds. pp.425- 
4,t7 
\[KIIIIO 73\] Kuno, S. (1\[)73). The st.ructure of the 
Japanese I,anguage. Cambridge, MIT Press 
\[l(uno 78\] 11uno,S.(1978). I)ai,w;t no lhmt)ou. 
'l'aishu kan, Tokyo 
\[Ohyc 7,r,\] Ohye,S.(1975). Nit, ieigo tie II ikakukenky u. 
Talshukan, Tokyo 
\[Saito !)2\] Saito,l{..(1992). Shinjou Jutugo mt Goy- 
ourouteki I~lmseki (l'ragtnatic Analysis 
about l~sychoh>gical Predicates). Nihon- 
gogaku, gol.l 1, No.6, pp.110- rIG 
\[Mikanfi 53\] Mika,ni,A.(tgaa). 
Gcudai (\]O\[IOIl JOSCl. t\[. Kuroshio-Shl,ptyan , 
Tokyo 
\[Minami74\] Minami,l,'.(197d). C, cndai Nihongo no 
Kouzou. 'l'aisht,kan, 'lbkyo 
\[l'ahne.r 86\] l'ahuer, F.IL,(1986). Mood and Modal- 
ity. C.2a.ml)ridge University lh'ess,Cambridge 
\[Sells 85\] Sells, I', (1985). I,ectures on Conl;eml)orary 
Syni.actic Theories. CSI,I St.~mford 
\['l'akubo 87\] Takubo,Y. (1!)87). 
Tougokouzou I<, I~utttttyakll.iotlhotl (Sy|ll;ac- 
t,ic StFIICI, IIFe }ttld (~onl;exl, m:d hd'ormation. 
N ihongogaku 1987--5,Meiji-shoin,'l?okyo 
\['l'e~,'a,m,ra 84\] 'l'e,'amur;t,ll.(1984). Nihongo no sin- 
takusu to imi II ';la.l)anese Syntax aud Se- 
mantics 11'. I(uroshio-Shuplmrt,'\['okyo 
\['l'eramm'a 90\] Teranm,:a, II .(1990). Nihongo no sin- 
l;ak,lsu t;o in,i 111 '.lapan0se Syntax attd Se- 
mmltics 111'. I(uroshio-Shuppan,Tokyo 
\[Tsuda 91\] Tsu(I;t, \[\].,llasida,K. ,Sirai,ll. 
(1989). JPSG l'arser or, (~onstraint, l,ogic 
I)rogramnling. 4th ACI, I,;m'ol)ean (2hal)ter 
\[Walker 90\] Walker,M.,M. li(la a.nd S. Cote(1990). 
(,'entering in .lal)anese 13iscourse. C()L- 
ING'90 
685 
