American Journal of Computational Linguistics 
Ell crof 1 the 41 
ANALYSIS OF JAPANESL SENTENCES 
BY US I NG ~EMANTI c AND CONTEXTUAL I NFORMAT 1 ON 
MAkBPO NAGAO AND JUN-XCHI TSUJJ I 
The organrantion of a nature1 langueg~ (Japanese) parser $8 d~brribed, 
The pslaes cw Transform fairly eoapl~s sentencrao into abtxtrau'lt s tru'rur as urked 
it 
for coae. A variation on the sys+em developed by Woads called an 
Aupntsd 
Transitio~l Netwo~k" is used as tRe proaxam for anelvsis. 
Tho parser utiliscb 
detailed semantic dictionary descsipedcans and contextual infamuat ion ~hst rac ted 
from sentences aqalvzad in sequence. It is clninncd that ineuirivc reRtsonlnp, 
which is nut easilu formdieed hv rigid lagiaal aperatiena, @lass an fmpartent 
role in language understanding Soor intuit i\"elv appealin8 scheous cf represen- 
tation for both the semantic descrihians of wolds and context are discussed 
1 
tkanlngs ~f verbs are described by using a case' concept, Additional ~nkorms- 
tisn is attached ta case frames of each verb te, indicate what changes the \case 
elements in the frahve my undergo and w'lat events mav occur in suqcession, 
bnings of nouns are also exqnessed in cdse-frame-like dcssriptions. Notins 
alscl kgve relational slots which must be filled IR b other GJO~~S of phrases, 
The context Is repregented in a form similar to that uf tke semantic n~twork of 
S-tm~ns sr the nodespare of Norwn along with saw dded special. 1ic;rs (KS - 
Noun Stack, HNS - Hvpotnetisal Noun Stbck, TL - Trapping List) These lists 
I 
contain objects mentioned ir prevlous sentences or pendlng problems whlch ma 
be resolved by succeeding sentences The objects bin WS are ordered according to 
their degrees f importance In the contest. Several new techniques based on 
heuristically admlsslble operatlone are presented to analyze 1) comple and 
long noun nhrases 2) conju~rtive phrases 3) anaphosic expresslan and 
4) omitted words in phrases or sentences 
The resulrs ~f applying the parsing 
program to tNe sentences In a textbook of elementarv chemistry are also 
presented, 
ANALYSIS OF JAPMESE SENTENTES BY USING 
SE 
xN"SR~DU~~OW...*..*.*O...*.*.O..** 4 
II LEXIML DESCRIIJTIOWS OF WORDS, . . . ....l..... 9 
2.1 Now kecriptfon 9 
2.2 Verb hscriptisn 13 
I11 ANALYSIS OF NOUN PHRASE . I . . . . . . . . , . . . . . . -22 
3.1 Properties of a Noun Phrase '2 2 
3.2 Analysis sf a Nom Phrase 2 7 
3.3. &alysf% QE Conjunctiva phrases 32 
3.4 Analysis of a S~mple Sentence 38 
IV CONTEXTUAL ANALYSIS . ........t . . t . 46 
4,l Basic Approach to Contextual Analysis 46 
4.2 &mr~ Structure for Contextual Informti on 48 
4-3 E~td- sn caf Odtted Words 52 
4.4 Psocegsxprg of haaphoric Express ions 60 
td MALYSLS OF GQWLEX SENTENCES . I .*+. 65 
VL CONCLUSZQP . . . * ...... ... ... . . .71 
I INTRODUCTION 
Ia this paper we describe the ssgsnitation of the natural language 
passes dewloped over the last twa years. T11ia fotms an important paet uf 
s question-answering system under de~@logmnt with natural language 
(Japanese) input. T11c patsar can transform frnir-tv cumplczl, srrltepces into 
abstract structures marked for case, It utilikes detailed srpbmntic 
dictionary descriptions and cont~xtual infurmatton skutract~d from th~ 
preceding sentences. 
For: the present, we heve confined the dewin of the system to the 
field of elamentary ehemlistrj where we can describe the semantic world in 
rather concrete terms At the same ti=, various compleu- events occur ln 
this field For example, substances which partleipate in part~cular events 
may disappear, new substances may emerge, or some properties of the sub- 
stance? may be altered To treak these comple situations, it was necessarv 
to formally represent relationships between events and cl~aages of state and 
to devise an appropriate schema of representing context, 
In mst appr~ache~ to the understanding of natural barlgunpe tl~roupk 
artificial intelligence, schemas which entail rigid logical operations are 
used to represent both knowledge and context Loglcal operatlons appear 
to be necessary for solving some kinds of problems in natural language, 
especially at the deep deductive level of unders tandlng. However, lntui tlwe 
reasoning is not easily formalized in terms of loglcal operatlons 
It is 
our contention that intuitive reasoning 1s completely based on the language 
activity in the human brain Associative functions relating to semantic 
similarities between words, semantic depth of an interpretation and 
probability of associative occurrence of events are inherent factors in 
intuitive understandinn and the seasoning process, 
Y. WII~ES (1975) in his eykatem carries out intuitive r~tarsoning by 
employing the notion of 'eemntie preference'. His system seems to work 
well an anralytfng local relationeahips among words, Mowever, in order to 
analyze amre global relationohips (e.lg., in dealing with complex cams of 
azlaphora) we require access to we information than can be contained in 
formulae (templates) a ed with the haxicon. We find Wilks' uss of 
'6s-inference rules' sahkar awkward. The system w~uld be much improved if 
accoarpanied by afi appropriate achema for representing context. 
v Case gsam~r' ~entence-analysis theories such as those of Fillmore 
(1968) and Celce-Muscia (1972) are based on the semantic relat ibnskips 
between verbs ad nouns -- events and concepts, R. F. Simmaas (1973; 19751, 
n (3973), D, E. Rumelhart (1973) and so on follow these theories 
to represent knowledge and context in their systems. We also adopted case 
s and nsodifded it to account for Japanese sentences. We represent 
cantext fn the form of a @@antic network. An input sentence is transformed 
into a c~rxeeponding deep saee etswcture, This structure is assillllsilated 
with the ~e tie network constructed from previous sentences, 
Japanese is a typical SQV language. The word order 1s rather 
srLItrary except that the main verb comes lest. Cases such as subjective, 
~bjective and dative as syntactically indicated by postpos~tions, but a 
petposition can be used for several deep cases anibiguously Hence the 
determination of underlying senteatial structures rests heavily on an 
undarstarading of the semantic relatione between the main verb and nouns 
kloreover in Japanese the words which are essential in understanding a 
sentence are often omitted without pronqminal reflexes. 
Our system can 
- 5 - 
infer from the semantic ddscriptions of wards what kinds of ghzsses should 
be supplied to fill lexieal gaps ?nd sgarch the contextual tcpreaentation 
to find appropriate fillers. 
The final analysis produced by our pamor ia a srtwntic netwvtk, 
This could he used fox the internal ,raps@sentatisn of data in a qupgtt~ft- 
answering system ox as an interniwdiete e~pr~ssiun in '61:mchin~ txansl,stiurt 
H~wu\!r, it is still too early to report on th~. rcsults of these ~ISL~~U~C~CIJI 
applications. 
The paxscr cottsists essrntisll wf four fixed cornp~>pcArs* 
I$ The grammar consists of rules written in PLATDE;, PLATOK 1s a 
new programming language which is a variant of the sqqtem deuelwped by 
Woods (1970) called 'Augmented Tkansitlon Network PLATON has add~tlonal 
facilities for pat tern matching and flexible backtr ackinp, A grammatical 
rule in PIATOW consists of two parts pattern rewrite which 1s expressed 
as a pair of syntactic patterns, and semantic and contextual check bhich is 
an arbitrary LISP function. Ffllen a rule is to be applied, the semantic and 
contehtual chech is emplnved to determine whether the rule is mwnr ically 
and csnteutuallv feasihlo, For the present we have about two hundred rules 
for the analysis of Japanese sentences, These rules are devised to combine 
Various syntactic patterns in Japanese with appropriate sewntlc and contevt 
ual checklng functlohs, PLATON is presented in more detall In another 
papec by Nagao and Tsujil (1976). 
2) In the dictionary are stored words along wlth their various 
semantlc relatlonshlps. We express the meaning of a word in terms of hov 
it may be related to other words, The meapinq of a verb 1s described In 
1 
the form of activity patterns' in the verb dictlonarv An actlt~itv pattern 
-6- 
is actually the case frame of a verb and additional related information. The 
case fr 
regresents what case relations the activity entails and what kind 
sf referents will be appropriate for each case slot, Adartions1 information 
provided feeds into the 'change' or 'causative' compahent used by Norman 
(1973). 
Such infortnstfdn indicates how one activity pattern my be related 
TO another by causal re, 
onghipe and what related change may occur in 
the eeatlantie network representing cbntext, From auch inflarmtion we can 
Infer what activitiefi and changt! will. foll~-)w the pre~enfr ~ctlvity, 
The meanings of nouns are also expressed in the caee-frame-like 
description8 They also haurn relational slots which will be filled in by 
sther words or phrases, 
3) The ~onterrtual representation is similar in fom to the semantic 
netw~sk of R. F, Simmons (1973;1975) or the nodcspase sf B, A, Nomn (1973). 
In tk~s representation there are two kinds of nodes. The C-node esrsespands 
to a concept typically expressed by a noun The S-node correqponds to an 
event. h event is .rs realization of an actian pattern and each argument of 
the pattern is assigned a C-node. C-nsdea are related to S-nodes by the 
case labled rclationa. These ~latbons are bidirectional 
The fallawing kist shaws the relations used in the network: 
(1 Deep Cue Relations. ACT', OBJ, PUCE, TIHE -- A deep case 
relation connects an S-node with its argument C-node, 
(ii) Attributive Relations: VOLUME, (r;OLOR, MASS, SHAPE -- An 
attribute relation connects a C-node with its value. We Gan dlstingulsh 
two C-nodes associated with the same lertical entry but different values of 
attributes. 
(iii) Taken substitution. TBK -- TBK is used to connect a node with 
-7- 
a lexical entry, 
(iv) Went-krent Relation MUSE, IFPLY -- Two S-nodes arc somtiws 
connected by n particular relation. The rclntions arc roaatims cxpscqst*d 
explicitly in the ~urfsjce sentenre hv a spocfal romjtmrtl~n sue11 aa 
NOBE (because), NA (if) and sa OR* 
In Our system the somantic network is acco~anicd with spt.cia1 Lists 
(Noun Stack-NS, Hvpathatirel Noun St,~ck-HSS, Trdpping L~z-TLL 
Ve call 
thqsc li~ta intrmrdiate Term bbmxy. Contcxtnal fu~~ctiuns votk on these 
Sis ts to seaxck appropriate nudes of the semantic ne twark correseend 
to the referents sf snaphosic expressions or the unex~sessed rLewnts of 
sentences, 
(4) 
Selaant~c and Contextual functions are propmmned in LISP 
These functions are lncorposated in the PLATON rules along with rewrltlng 
patterns A contextual function takes as arguments the seaantls constsalnts 
a target node must satisfy and returns the node when an appropriate node rs 
found from the semantic network A senwntic fu~ction cheeks dcscriptiuns in 
tt~c dic ~~QII,ILY te7 dc turnline ~~hcrt~er tEw cwrnhin,~t ion of two 'r~ards is 
sen~anticallv pedwsibl~, For analp zing nourl-nc7tnl combir~ations , wc provide 
sixteen semantxc functions. 
The whole system is wrltten in LISPl.6 whlch 1s implemented on 
a minicomputer, TOSBAC-GO. The mlnlcomputer 1s equlpped wlth 64KB as main 
memory and 512KB as secondary core memory The LISP1 6 uses the secondarv 
memory as virtual storage The drctlonary conslsts of about 400 nouns, 
200 verbs and lLQO adjectives and other categosles The parser with rules 
and the dictionary occupies about 50K cells The LISP uses a soft-ware 
-8- 
paging mechanism, and the main memory is rather smell in comparison with the 
secondary mmory The content of the dictionary is stored on a disc in 
the form of S-expressions. 
Consequently, the speed of execution is slow. 
It takes typically about 10 to 15 minutes to analyze sentences which oontain 
3 to 5 simple sentences and 10 to 15 noun phrases. 
XI LEXIUL DFSCRlPTIONS OF WOW 
2,-1 Now kscripkiern 
&st nouns have a definite meaning by themselves. We call these 
Entity Wows. kn entity noun is considered to represent a set of objects, 
and therefore is taken aa a name of the set. The sbjects belonging tta the 
set $hare the same properties By lntsoducing another property the set 
my be divided Into a number of subsets, each of which is expressed by 
another noun, We describe such set-inclusion relationshrps and set 
propestles ln the noun dletfonaq. 
We represent a property of a noun by an attrxb-ute-value pair 
expres~ed aa (A Vb . For instance, the dirtinner entries for the nouns 
'wtcsiel' and ' liquid' are : 
wterJl~aI ( ( SP> (AmR(STATE) (MASS) (COLOR) (%WE) --------- 1) 
liquid ( ( SP material) (ATTR (STATE *LIQVID) (WE NIL )) ) 
Re descriptions (STATE) (WSS) and so on m the deflnltlon of 'materlall , 
I 
Lack values (V) showing chat material' may have arbitrary values of these 
zttributes, In the definrtlon of 'Ilquid', there 1s a $P-link to 'raaterlal', 
which meam that 'material' 1s a super-set concept of 'Ilquld', or that 
1 
' liquid' is a subset or a lower concept of materxal' . Objects belonging to 
-9- 
s subset are considered to have the earn properties 18 the abjeca laf tha 
super-oat, in addition to the proprties described axglLcitLy in its 
dafinition. 
By the above daecrhpticor, rn ran sea that %k@ value bf the at tr%bure 
STATE af 'liquid' Ler *LIqWSB, and that u8f %WE LB th~ I~QI%L~$ V~~LUQ 85L 
*SYtla *LLQWXD he ona of rha pridtiws vrllua mrkrrs, Thr prleieim voZur 
mrkers are indicatad thl~ the preceding J"na value NIL indfcstab that 
'liquid' can not have any value of SHWE iv trrcin~ up thc SF-links, uc ran 
rstrlew all. the (A V) pairs of an o'oj@wt. We assum thc value of an attri- 
bute of a lewas coneept has precedence over that of eha upper concept Far 
inutmca we csrl obtain the following full, description of ' liquid' 
liquid ( (ATTR (STATE ELIQUID) (WE NIL) (INSS) (COLOR) ------ 11 
These upperlower relatianships smng entity nouns cr$ not expressed 
by a tree structure SQ~ nsuns may share proparties with more than one noun 
'Water' is such an example -nter1 has the properties OF both 'liquid' and 
1 compound' 
Since wc permit n noun to have several upper concepts, the sela- 
tiarlships arc represented bv a lattice ,%s shown in Figute 2 1. 
chlorlde 
(salt) 
solution copper 
- 
zinc 
7- 
salt liquid 
solutioo amlm~nia 
-. 
Frgure 2 1 Upper-lower relationshrps 
dmeng nouns 
Although ast nouns are regarded ae entity nouns, there are a few 
nomg which hawe relational function@. Mte call them Relational Noma; 
'Fathart ka r famili&r a pie* In order to identify a pesaon indicated by 
thc word, mi hrvt EO haw whose father he ie. 
In the chemical field we ran 
auily find auch nome (e. 8. , 
'weight' , ' temperature' , ' color' , qnnd 'mess ' ) 
*E"hrdss &re erJlcd Attribute Wslane. Their maninge are deacrilaed in a different 
way from that of ordinary nowe, Figure 2.2 shews some @,xempliea. tiare, A-ST 
draipater, the s tmdsrd attributive rcldtion which is expressed by the word. 
The description (t4F N-A) shows the noun belongs to the group 0f attrlbuto 
( WRISA ( (WP W-A) (A-ST WLW kWSS LENGTH A 1 
s LZB (SP ZQKUSEZ RY8 b>) 
attribute quantity 
( IRO ( (NF M-A) (A-ST GOL~~R) (SP ZOKUSEL SHITSU 1)) 
colar attribute qua1 ity 
* 
An attribute nsun may express amre than one standard strtibute. WKISh 
(size) expresees VQL MASS, LEHG;TM QR AREA. The attsibutc it expreoqes 
in csntext depends upsn what entity noun is used dth it. 
aft Attribute noun8 am further ~Zaseif ied inta twa groups, quantitative 
and qualitative A guc*aitati~g attribute noun cannot be a case element af 
a verb wtaech regwfres guntitntive nouns. Tna verbs FUERU [increase) and 
MERU (deciease) are such &xamples gf verbs. 
Figure 2 2 Attribute Nouns 
'Liquid' is another relational noun. The Japanese word whirh 
csrreapoads to 'liqula is EKITAI. While 'liquid' in English can be elther 
a now or an adjective, ERITWI in Japanese is ca~rgorrzed synta~t~calJy as a 
nom. But @emantically EKLTAF has two diffe~ent meanings, one corresponding 
to the liom usage of 'liquid', the other corresponding to the adjective 
usage of it. The noun EKITAI in the adjective uasp is collsd a Value Noun 
wf th the! attribute STATE. Anofekes word IRB (scd color) 18 alga a vrlua 
noun of the attsibuta COLOR. Figure 2.3 ahws rlrc da~cription of th~n~ 
nouns in the noun distionerv, 
( EEFXTAT ( (NF N-E) (SP BUSSHITSU, (ATR (STATE trqvrs;rs CSWPE ~11;) 5 
liquid wt@1'f;;laX 
( (NF N-V) (V-DESCKZPTIUN (STATE &I;XQVI'I?) 1 ) 
There are other kinds of relational nouns Action, Prepositional, 
hnphorie, and Function nouns, h action noun 1s the nsmnalization of ,I 
verb. For example, KANSATSU (obserwat~on) is the nodnallrzation wf the ~erb 
Q'iNSATSU-SlrRU (~bscmel 'bu'c describe this in the dicti~nax~ kv giving a 
link to the ~xigit~~3.l verb and bv adding other infua;mtS~?n 
There arc nut t positions1 pb%rtic9cs Ln Japanese fur r.wr\ 
preposition in English, Sow special nouns plg the sole af English 
preposltlons. We call such nouns Preposltlanal Nouns Because a prepasl- 
tl-onal noun usually has more than one meaning just as an Engl~sh preposl- 
tlon has, we attach semantic conditions to help disambiguate then Flgure 
2.4 shows example of leslcal descriptions of prepositional nouns 
C 
before 
in %mnt (3% 
(\MD (WT W-E) (LOWER DOUGU 
instrument 
( (WP W-P) (F-DESCRIPTION 
Figure 2+4 Prepositional Nouns 
Corresponding to each waning we give a triplet The Elrst element is the 
aemntic condition, If the ~ondition is sutlsfled, the corresponding second 
elewnt 28 adopeed as the waning, If not, the next triplet is tried The 
q~sond clcmnr of B triplet represents thc wt~ole waning of the phrase. For 
exrqle, rb *ole maning of the phrase TStKrtiUE (desk-entit ncun] NO [of1 
Ur: [an-prepositional noun] (on the desk) 1s PLACE The third element of a 
rsiplae expregnoe she reletionstlip hk which the other notln ~n the phrase 
wy ~w~iajibze eke whole maning. 
2,2 Verb kacriptiora 
Verbs, adjectives, and prepos~tions in Engllsh have relational 
-wing@ wlth nouns. A verb represents a certaln acti~ity, Fih~le the agent 
mooc~eted wzth the activity is not ~nherent to the meaning of the verb 
(neither is the object the activity affects, nor the other components) 
These components appear in a sentence with certarn loose selatlons to a verb 
- 13 - 
In our system the meaning QE a verb is d~~csibed by settrirzg up k;emsal 
relational slots whLch will be fillad in by nouns, JR this sen&@ the 
waling of e wcrb is not confined to itself, but tw relrtgd to mma. 
We describa rhase relations by u~ing the cone caneept intr~dured by 
6. J. FIllwre (1968) Case GZBY be Iiaohed UPCI~ 18 a mla which an rbjrct 
pknys in sn actlvitv, Because scwllraJ. abject& ~,uuallu pertizipare in an 
activity, them are s~varal easeas aasncistrigd wit11 an aitrfvit%. kl nbfcget 
ie e&$raascd b) a twun phrage, and wn activity hv a wrh, A sentence 
instantiates afi activity b) supplving noun phrases to the cases associated 
with the activity Me call such instantiated activity LEC~ Et*~ntr The 
prslskem is tea decide what case a noun phrase holds in selatisn ts a verb in 
an\ particular event. 
Tbu& there are usually some syntactic chess in a sentence as to 
how it instantiates an uctfv~ty, they are not enough to decide the ease 
rekarionships between nQu;n phrases and a verb. To establish these relatiom- 
skips we need both svntactic and sc~k~ntic information, A verb has its own 
sp,t.cinl. usage patterng, Z"t~ct is, certain caws ate ngcewsaxy for the 
activity and certain objects nx+e prclernhla 8s fillare fur the case. Q@ 
aall these labled patterns Case Frames for Verbs, and express them ss a list 
of case palrs such as (WE NOW) A verb usually has are than one case 
frame corresponding to different usages, A typical description of a verb 
is shown In Figure 2.5 
CCF 
( ( ACT MINGEN ) ( OB J KOTAI) ( IN EKITqI) ) 
human being solid liquid 
(( ACT NINGM ) ( OB J KOTAI) f l VST) ) 
human being solid 
(( ACT SAN ) ODB JKINZOKU )) 
acid metal 
Figure 2.5 A typical dpsctiption uE 
a verb 
According to this description, we underskand the verb TOKASU (melt, 
diassolve) hae two different usages. In one usage the verb takas thc ACTOR 
case, and prefers to take the sub-concepts of the noun NINGEN (human being 
taa the case element, In suck a way case f ramc descrip tians are closely 
tied to noun descriptions, espec~a1l.y with the upper-lower concept relation- 
ships amng nouns. 
Rere are twca types of cases, Intrinsic and t xtrfnsic cases, The 
iratrinsis caws st a verb ate eesantial ones for the ~ctivhtv, but extril.lsie: 
caw@ am not. For exampl~, the cases of TIME and PUCE, which express 
when and where an event occnrs, are extrinsic f~r ordinary verbs, kWst 
aetbvitxes can be modxfied by these extrlnslc cases, bur the kinds of nouns 
preferred for these case elements do not s tsongly depend on the kinds of 
activities* Therefore we descrrbe only the lntsinslc cases In the verb 
dicrj.onaky. We set up fourteen cases as shown in Table 2.1 for the 
analysts of sentences In a textbook of elementary che&str]i 
TABLE 2 ,1 
(1) ACT : ACTss 5s responsible for action. 
IOU -43 SHTK -HI t RERltI 
he- (ACT) 
?111R - 
sulk ur -OW test tube-(IN, PLACE, etc,>put in 
He putq sulfur in a test tube 
- 
In the chemical field, o chemical object is aFtrn scpitdtd ss ACTnr 
of ,an action, thouah it dors not ~xcrcise i$ten,eiun in lrpetd to actionmJ ror 
OX~IP~~ the unde~lincd ward in the follc?w?kR erntTncr is rr~ardad aa ACT. 
-WA 
- 
WU -0 TOUSW, 
1ci3 -(+BJJ- cases) e.uppet- -0BJ melt 
- -- 
wsochloris - acid melts copper, 
(2) SUBJ : SUBJect is the primary topic of a sentence 
(a) KITAI-NO TAZSEXI -CA FUERU . 
gas volume -ti'C~$ increase 
The volume sf th@ gas incrwses 
-- 
Cb) IOU 
- 
-MA 
- 
KIIROI 
sulfur 
-- 
-(SUB~:, 
--- 
vellow 
(3) OBJ OOBJect is the ~ecci d11g end of an octiwitv. It s affected by 
the activity, 
(a) IQiRE-GA MIZU -0 - NESSLhRU 
he-fS1M) water - (OBJ) - heat 
He heats the water. 
I__- 
(b) TNSAN - GA AEN -0 
- - TOUSU 
ydrochloric acid -(ACT) zinc -(OBJ) -- me1 t 
Hydrochloric acid pelts zi nc . 
(4) IOBJ This case 1s semantically the mast neutral case, It is an 
object or concept which is affecte? by sn actlvlty, and which 
is not OBJect, This case is usuallv specialized by the other 
cases such as PUCE-, TO, IN and so on, depending an the semantic 
interpretation of the verb ltself 
TABLE 2.1 (continued) 
(4 -0 - ENSAN -NI - 
copper -(OBJ) .I 
:.EN) 
(~ocrrcne) dips copper - in 
- acid 
(5) 
FRMM , FROM describes e former position or state in time or space of the 
entailed S'IIBJQL~ or OBJccF of the verb. 
(a) 
BEEM -E EK JTAI-O trrSUSU, 
) beaker -(PLACE) liquid-(QBJ) pour 
(So~one) pours the liquid TC from the -11411p test tube into the beaker. 
(6) 
RESULT ; RESULT is to the future ss FRW is to the past It describes 
the resultant position or state 4s the entailed SUBJcct or 
OBJect of the verb. 
(1 HZU -GA - SUIJOUKI -MI - UQV, 
water -(SUBJ) ,$team -(RESULT) becw 
The water becomes steam 
(7) INST ; IWSTrumant is an object used as the tool or device by which sn 
aerfwlcg is carried ~ut 
asra-pasar 
(a) -- 
-BE 
- - 
MTZU -0 WSSmW. 
-(INST) - water - (P8.l) heat 
(Solpeonr) heats water ~JJ - a gas burner. 
(8) TO ; Thi~ i.8 the destination in tiw or space of something in the action 
(a) s~BW -GA ARU TOKI 
7 
MADE NESSHI TSUZUKERU. 
water -(SUBJ) be gone -- time-(T0)till heat con t inwe 
(Somoqe) csntrnues to heat (~t) t~ll the water is gone. 
TABLE 2.1 continbed 
(9) FAC3 ; FACT is used to &dieere sentential cotnplrnuenti~@ra. 
(a> KOBEe -0 SMSTS~YOwOZUW-NO 
-- ------ eNlsurrv -ro =...a,* et., 
it -fBBJ) - the conservation sf m~s 
1I_Y__--**YIPI1. - by- 
a 8%~ 
*m -(PACf) ---a- &*all 
Fk call it the law of conseruatiwn of mass, 
--- J____=__EI___ilP 
(18) PUCE ; PLACE is used to indicate lecatiuns in spare of the action 
('3) ARKWRU-MPt,t=NO YOKO -NI 
- 
BE -b QKL'. 
alcohol lamp sidr - - -(PUCE) 
-- beaker -\OBJ) Put 
(Sowone) puts a beaker ---- on the side of ah alcohol lamp. 
(11) IN , IN indicates a mre specific relation tp PLACE. 
(a) bIIZU -0 SHIHNW -NI - IRERU , 
water -(OBJ) tube --(IN) - pour 
(Sonueone)p~-urs water ---- in a test tube. 
(12) SOURCE ; This shows constituent materials of compounds, 
(a) ~SOSWATORILPKT -Wd ENSU, SANSC), &iTORXLPn! -KARA 
- 
sodium chluratc -(StIBJ\ chlerfnc 
Sodium chlorate conslsts of chlorine oxygen atld sodium. 
(13) CAUSE , This shows a reason or cause of the actlwlty. 
(a) WESSHITA-TAME -NI 
- - HAGESWIKU KAGOUSURU 
heat -reason -(CAUSE) violently react 
Because (someone) heats (them), (tney) react vlolently. 
(14) TIPIE , Time indicates location in time of the action, 
(a) NESSHITA -TOKT -NI 
- 
SANS0 -GA HASSEISURU 
heat 
- 
ti= TI) owgen -(SUBJ) be generated 
Oxygen is generated when (someone) heats (it) . 
- 18 - 
In .order to rq~olve the cayntactic ambiguity of a sentence, it is also 
nscaoaery ts utilize Contextual Snformtion obtained from preceding sentences. 
hen one knowa a certain event has occurred, he can anticipate succe~sive 
events that will occur and what changee the objects participating in the 
evant will undergo. 
Thie kind of expectation plays an important role in 
underetanding eentences. Various kinds of associations clucter conceptually 
arsund inddvfdumb activities, One can perform eontcxtunl analysis of 
language by explicating these associations, 
We append this kind of experiential knowledge t/b the case frau~s of 
vetbe The follawing two items are described for each vcib in the verb 
dictionary: 
(1) CON thie refers to the conseqdent activities which are likely to 
Eollow the activity sf the verb, but not necessarily, 
(2) 
S this refers to the resultant effects on objects in view of 
how the objects are influenced bv the activity. In our system 
the influenre on the objects is described by the following 
three expressions: 
(a) 
( ADD case a-set-of-(A V)-pairs ) 
(b) 
( DELETE case a-set-of-attributes ) 
(c) ( CREATE lexical-nam-of-an-objec t a-se t-of- (A Vj -pairs) 
(a) mans that the object rn the case indicated by the second element comes 
to have a set of prgpertles lndrcated by the thasd element. (b) is for the 
deletion of a set of properties from the object. (r) shows that some objects 
will be created by the activity. 
A typical zxample uslng a CON expression is shown In Figure 2 6 
t -- 
r A 
'j 
CI 
4 
pa 
d 
k 
t 
D 
3 
U 
4J 
(Ti 
QJ 
u 
Qi 
s 
0 
C 
-4 
'J; 
ri 
k 
3 
w 
4 
3 
fV2 
G1 
C 
4-4 
- 
ul 
11 
0 
4 
3 
[I) 
IU 
k 
CU 
s 
C, 
-a 
c 
a 
h 
Z 
0 
U 
w 
0 
In this expression one can see the verb TOKASU has two differetlt meanings. 
One cottesponds to 'melt', and the ather to 'dissolve in'. 
When we analyze 
the aentense, 
MU -6 
TQKASU . 
c~gper- (0B J) dissolve, melt 
Sawone mlts copper 
WP adwt the firgt enae fraw of TOKhSU (melt1 because it gives the 
highest mtched value against the sentence (aee section 3.4). As the result 
~f evaluating the MWS exgrefssion in the case frame, we conclude the copper 
1 
is nm in the liquid state. In the lexical description copper' is a lower 
concept af 'solid', so that copper in general behaves as e solid object. 
But the copper in the above sentence comes to have the attribute value 
pair (STATE *LIQUID) and will behave as 'liquid' in the succeeding sentences. 
On the contrary, when we analyze the sentence 
SNIQ -0 MZZU -MI TOUSU 
aalt -(OBJ) water -(IN, PLACE, etc,) melt, dissolw 
Soeane diswalves salt in water. 
the second case frame of TOKASU (dissolve in) gives the highest matched 
value After the sentence instantlate9 the case frame, a new object (l,e*, 
a solution which conslsts of salt and water) wlll be created. 
CON AND iYl'MTS are thus important in the contextual analysls of 
sentences. The detailed analysis procedure using these expressions iS 
described In sectron 4.2. 
rIr MALYsrs OF NOUN PHRASE 
3.1 Properties of a Noun Phrase 
In Japanese, two or mote nouns are often conentcnnted by rhc 
pastpaeftion NO ta form a noun phress, Because there are any dhff~tent 
sewantie relationships smng nauna concaten&evd by NO, we nust Secfdr kkat 
relationshipe may hold among the nualns. Qpical caamplcs arc shoty% in 
Figure 3.1, 
EhitfTAJ: JOLTAS. -NO SMSO -8 ThZSEkl 
aiquid a tate Q= gem VC.~UW 
the volu.wac) of the oxygen in the st&@ of liquid 
M@U -NO AT -NO ~vATORIL?!IJ -NO TAISEkl -NO IJE2K.A 
react ion after sodium 1 UIB~ change 
changes of the sodium's vulume aftel the reaction 
Eigurc 3,l. Examples of NOLWNO phrases 
Thc phrase NOLWNO can mdift , in principle, anJ or ,111, of the suc~e~dbny 
modification relationships are syntarticallv permitted, frde must decide which 
one is correct by considering semantic restrictions. 
We have identified slxteen semantically acceptable NOUN NO NOW 
combinations. These are shown in Table 3.1. (See pgs. 23-25 for th~s table ) 
Corresponding to these relationships we prepared sixteen pslmitlve functions. 
These functions are applied in turn to a noun phrase to declde what relat~on- 
ship holds between two nouns, The order in which these functions are applied 
1s based on the frequencj and the tightness of the relations, Each function 
checks only one semantic relation Xn order to .illustrate how these functions 
perform their tasks, the following exaxiple of noun + prepositional noun' 
- 22 - 
TABLE 3 1. Admissible noun-noun 
Combinations 
(a) ( value ngun )+( attribute noun) 
(ex) KQTAI -NO JQUTAT 
solid atate 
(2) ( value noun)+( epP3.t~ noun) 
(ex) EKITAI NO ZQU 
Iiq~~icf sulfur 
(3) ( entity noun)+( attribute noun) 
(ex) EKITAI -NO IRO 
liquid calsr 
(4) ( noun)+( psepoei tionial noun) 
(ex) HMOU -NO WE 
reaction before 
(ex) MOT0 -NO BUSSHXTSU 
former material 
*In tkdb usage, the attribute noun should Be 
mdlfied by another noun or adjective, whrch speci- 
fies the value sf the attribute. 
(7) ( now)+( actiotr naun) 
(ex) SMSC43614 -NO WGEN 
oxidized deoxidizati on 
copper 
ZRO -WO wmKa 
color change 
TABLE 3,l continued 
(8) ( time)+( noun) 
( (SHIKEWU -NO) MhU - FKIFAIh 
test tuhc in liquid 
t 
?The noun-nuun rt~mbinatiw, test tube-SL7 
in' expresses the 'place1 in the test tube, 
(10) ( noun)+( conjunction noun) 
(ex) SMU -NQ TME* 
oxidi~atlun in order ta 
b\t leerason of 
Japanese, so= nouns are used te elucl- 
d,3tc the cb3sc relationships hr tween s nQwn phrase 
and 3 verb TI72 noun "SIE ;tn this c\ -qplr. 
expresses ck;lslo.; such 3% CAL'SE lsr P11IPOSE. 
The Elrst entltj noun 1s a constltuent 
element of the object uypressed 'try the second noun 
(12) ( entlty nounlS(entit\ noun) 
(ex) SANWIOU -NO SAlYSO* 
oxidized copper oxygen 
*The second noun rs a constltuent element 
of the object expressed bv the Elrst noun 
TABLE 3.1 continued 
(13) ( entity now)+( entity noun) 
(ex) SH'3BKmM -E30 SOKO* 
tcet tube bettsm 
*The second now refers to part af 
the object expressed by the first noun. 
(14) ( entity nsaasrs)f( entity noun) 
(ex) URI HATWRI -NAM -NO KINZQKU* 
pstasalua, sodium ~tc*. metal 
1 
*The nouns 'potassium' and s~dium' 
are lower concept nouns of the last noun 
'mtaX' 
(15) ( na=)+( noun) 
(ex) SHITSmYQUHBZOW -NO NOWSQKU 
the conservation of mass law 
2 
(ex) lcrn ATMI -NO CW1UR.A 
per lcm2 pressure 
phrase is ginn. 
The noun IYUE is a prepositional 11c7wB and its Rrwuslltic d~q~riptlon 
is shown in Figure 2,4. Wc. note thst this world has two JL t fetk~nt owanitrtgt* 
-NO tUi5 
t lnw plteceriit~$ 
17 lncc in front vf 
The fut~cttun far tklr m;;ldlpYis of this kit~Q of ~II~QSC c,.hec,'k% ,zt titst 
this function fails and laturn$ the value N3L. In this cndnplc, b~ca~~sv 
the word bL'2.E is a prepwsitiunczl noun, the checking proceeds furtIlt91 Tl2e 
description in Figtrle 2 4 shws that if the preceding noun 1s an action 
noun (i e, , if it is a nodnallzatlon of a ~erb) then IkE has the f~rst 
maning. Recause the noun JIWX (e vperiment) tjatisf ics th~s condition, 
the checking succeeds and the function returns the value T The result 
of the annlvsis is skoim in Figure 3.2 (a\ , thc other hand, iE the 
input is 
TStKlT -Nc3 lL\E 
desk before, in front of 
then the word TSUhUE (desk) satisf~es the condithon of the second meaning, 
and the result is as shorn In F:igure 3 2(b$ , 
(a: JIMKEM -NO 
experiwnt before : time 
In front of : glace 
(b) TSWUE -NO 
desk before 
place 
Figure 3 2. Results of analy~es of noun and 
prepssftdsnal noun phrase 
In thls ray the sixteen checking functions not onl? test whether a 
certain sewntic selat~onshfp h~lds awng input w~rds, but also disambiguatts 
ehg meaning4 of input wsrds 
Wc analsrc? a maan phraec bv using the above sixteen rheck~ng 
function@ eubject to the linltation that selzied noun gceups aav nc?t overlap 
1 
stated before, noun + pas tposltion NO' phrases and ad.iectives can modify 
only the eucceeding nouns. We stac~ in the temporary aack noun phrases and 
adjectivee for which the nouns to be modified have not been deteruned, The 
analysis of a aorta phrase IS carr;ied out by scanning words one-by-one from 
left to r:ght, If we scan an adjective or a dererrmner, we stack the# word in 
the temporary stack. If we scan a noun, we p~ck up a word from the tempciar 
stack and check whether: it can modify the noun. mi8 checking 1s done by the 
above f~rfons if the stack word is a now. Me 8190 have the checking 
functions mlating nouns to adjectives ot detedners "a"r;le dictimary, 
cantent: tsE arr adjkzrtiwe is brnt the sew a@ that ut a valt~e The 
gewntic checking funrtlvn hetwccn &III adjective and w nebv vhbl teat whetbar 
the noun can have the attribute which is mdftfkakl~ by the adjetti=. The 
chocking of the determiner differs ~naarcawhnt and is expleincd in a Isrer 
chaptor 
Thv o~PIEIIII~ ~KQCCPI wikl stop wllrtl there are nv words In the 
temporary stack or a wpJ is picked up that fails to w$ify the noun bring 
gcanncd, The now1 i$ then stacked in the temputav stack, If the ee 
stack contarns only one noun and there are no words €a be scanned in the 
noun phrase, the analksis succeeds ad returns the noun In the stack me 
returned noun is called the Mead Noun of the noun pt~sase These processes 
are ~llustrated in Figure 3,3. (See pgs 3-30 fur this figure 
If th~rc are no words to be scar~ned next and the t~mp~)r~r;~"fy st~ck 
contains than CYIC word, the11 tk~ anal\$ib fails and backtracks to the 
dcrisiun patr~t~ of thc program, h d~?ci?*lo'tl puint in thc ~xrtalvsis of a ni 
phrase is snv point at which two words haw brezr~ rolated semntLcral1.p. 
relationship between two words established during the analvsis is the? me 
determined by the function which succeeds first. Because the order of 
checking functions is somewhat arbitrary, in some cases a relatxmship Wcfi 
has not been checked may be preferable to the established relatiooship. lhis 
is illustrated in the examples below 
SHIKENW -NO NAKA -NO AKAIRO -NO EKITAI 
test tube in red liquid 
(I) Tewsrary Stack = empty 
test tube -NO in -NU rad -NO liquid 
7. 
I 
scanned word 
teat tube -NO in -NO red -NO liquid 
t. 
~kanncd word 
&A check of the ~erpdntic relationship between 'test tube' and 'in' is perfornled 
**The phrase ' test tube -NO in' is transformed into the f~mi PUCE 
0 
test tube -WQ in -WQ red -NQ liquid 
f 
scanned wordQ 
*A check of the semantic relationship between 'place' and ' red' is performed, 
but .it failed to establish o new concept. Therefore, 'red' is placed on the 
top of TS* 
test tube -NO in -NO r~d -WB 
scanned word 
*The next scanned word IS 'liquid', Since it is a nourf, a check of the 
relationship between the noun and the wsods in TS is perforcxed. The check 
succeed8 because the combinations (value noun)+(en tity noun) and (PLACE)+ 
(entity noun] are semantically permissible. 
test tube -NO in -NO red -NO llquid 
Figure 3 3 
fiere are no words to be scanne'd, and the TS contains snlj* one word, Hence, 
the analysis of this noun phrase succeeds. 
Ihe result is as follows. (The hccld noun of thi.; noun phrea~ is '1 Lquid' * \ 
Figure 3,3 continued 
EKITAI -NO JOU"I-'AZ -NO 
liquid state 
the ehanae sf state of the .liquid 
HEW 
change 
sxysen in the liquid ~tate 
Jn the ldr~t examle the ward Jf)UTAT (state) designates an atttibute 
of EKITAI (liquid) end EKITAI co~responds ta a visible, teal object . JDUTAl 
(ststel in the second example dig'lgnatea en attribute of SANS0 (oxygen), and 
the w~rd EKITAl doee not correspond to a real object but is used to specify 
the attribute 'state' of the oxygen. hese examples show that the word 
EKZTAJ (liquid) has two different usages. According, to these usages, there 
are two dffbereqt eee~antic construeticans sf the phrase EKXTAI-NO JOWAI as 
shm in Figure 3.4. 
KTTAI (l&uid) -NO JOWAI (state) 
(1) 7 STATE 
liquid 
indicates 
an abject: 
Figure 3.4. Two different deep structures 
for the phrase EKITAI NO JOUTAf 
Because we analyze a noun phrase from left to right, we cannot 
determine +ich usage is correct unti,l we recognize the rightmost word HENKA 
(change, transltioo) or SANS0 (oxygen) . Hme~#er, a semantic checking function 
disadiguates the multiple meanings of the word EKITAI. If the disambiguation 
14 recognized to be .hcorrept in subsequent processing, we must be able to 
backtrack to the decision point at which this temporary disambiguation was 
made. We implemented such a process by using PIATON'S backtracking 
facilities. This pxocess is illustrated in Figure 7.5. (St*c pps. 33-34 
for this figure), 
3.3 Analvsis of Conjunctive Phsrnse~ 
1 
The words in Japanese which corraspor~d tu and' and 'or' arc 
and (closed lis tine) 
and (open listing) 
In Japancsc as wcll AS in English it is difficult to dt.tr.rtu;lnc. tilt. scope of 
a co~ijut~ctio~~, R~PP are som ~htdses whit11 h6~we the saw sbntdctic 
structure, hut semnticallv Jiffctcnt ronstxurtaons. Sonw cyamples drc 
shotm in Figure 3.6, On the other hand, scsw phrases have different surface 
structures but convey the sane meanlng as 1s Illustrated Ln Figure 3 7 As 
there are few syntactic clues in these examples, we must analyze them b) uskg 
semant:ic information. (See pgs 35 and 36, respectively, for these f:igures) 
At the first stage of the analysis of a noun phrase, we try to find 
conjunctive psstposit:ions If we cannot find them, the normal analis is 
sequence described above is applied on the noun phrase, If there is 3 
oonjunctiue postposition, the following steps are performd 
- 32 - 
Input ; EKXTPnI -NO JOWTAX -NO HEMA 
liquid state transition 
, Transition 
JmJEcT 
Input ; EKITAZ -NO JOWAL -NO SnNSO 
liquid state OT? l2en 
(i) TS = empty 
liquid -NO state -NO oxygen 
+ 
1 
scanned word 
liqurd -NO state -NO oxygen 
T' 
sei~ned word 
liquid -NO state -NO oxy en 
$ 
seadned word 
%t this point, the Elrst meaning of 'liquid' has been adopted because 
the checking fu;nctlon for (sntlty noun)+(attrlbute noun) 1s applied before 
the fwcti~n for (value nom)+(attribute noun), That is, the word 'liquid' 
indicates a physical object. 
**The semantic check between 'state' and 'oxygen' falls, because the 
attribute noun 'state' has been linked to the llquld by the relation 
UTR-ATR and an abtribute noun cannot be linked wlth two different entity nouns 
***So the program will go back to step (11) . 
Figure 3.5 Example of backtracking in the 
analysis of a noun phrase 
liquid -NO -N11 ~~v,?acn 
*The se~umtic chcck b~twecn ' liquid' ;and 'a~etr' procr.cds tin thet * 
Thc semantic chrehing function fo~ (vsluc naun)+(attiihutc noun) suc~ct. Jr;. 
This function adopts thr src0tld mnninp, of ' Ilquid' . 
*At this timc, because the noun 'state' is onl\ linked to the lalur 
1 
*LIQUID. the check between 'state' nnd on*genV succccds The result is as 
f 
follows, Notice that the WUI~ liquid' does not rwress a real object but 
the value of the attribute 'state' 
(1) RYWTX3U -0 IX3U -TO IOU 
copper sulfide capper (and) sulfur 
copper (and) oxygen 
(2) copper 
copper i(;lfide 
hvdragen oxygen ratio 
hydrogen beaker in liquid 
number 
I,,, 
(2) hydrogen 
\ 
PARA 
liquid 
PUCE 
aodius chloride solution (and) sodium chlorate 
(21 
I 
NATO;RI?DKI -NO TAISEKI -TO SHITSURYOU 
sodium chloride volume (and) mass 
1) solu tlon 
(2) volume 
SOLVENT 
/ 
UTR-ATR 
sodium chlorf de, sodium chlorate 
mass 
ZATR-ATR 
sodium chloride 
F.9.gup.r; 3.6 Examples of csn-junctive 
phrases 
- 35 - 
(2) SaZNSO -TO SUISO - N13 TAf SWT 
ox\*gen (and) hvdrngcn v 0 k t~~ 
(1 SANS0 -WO SHXTSLRJQU -TQ SUSO -80 TAXSEKI - 7'0) 
.e4uvgen mass (and) omgen 'L olume 
(2) SMSQ -NO SHLTS'LrRIOU -TQ TAISEKX - (2'0) 
rsmgen mass (and) vnlume 
*(TO) 1s an optional element in these sentences 
Figure 3 7 Examples of differing surface 
structures conveying the saw 
meanings 
1. The conjunctive postposition 10 is of ten followed by 
* 
another pos tposition TO in the succeeding part (Figure 3.7) . Hence if we 
find M in the phrese, we do the following; if not, go to step 2. We 
aearch for the wecond postpasition in the Bucceeding past. If it is found, 
then the noun phrase before the firet postposition and the noun phra~e 
interposed between the fi rvt and the second pog tpssitiona me paralleled. 
WQ eaoploy the norm1 noun phrese anaiysia to the interposed noun phyarae, ther 
Rn to etey 4, TE w cannut find the wecod pnstpoeitkun, we then 80 to st.ep 2 
2. If a sonjwncthw postposition i~ not TO, or there is no 
eecend Ta, we execute the follawing sbbsteps. (Noun-l de~ignates the noun 
befom the firax postposition .) 
a. Searefi far a noun   den tical to Noun-J- ir the suceexding 
part. If Eomd, let it be Noun-2, and ga to step If 
b. Xf Nowt-h rs not an entity noun, then search for a noun 
which belongs to the same categorv as Noun-1. If found, 
let it be Noun-2, and go to step 3 
c?, Search far a noun which hiis an upper concept in swmn 
with Naun-h. If found, let it be Noun-2, and go to step 3. 
St9 1. The phrase between the postposition and Noun-: are analyzed 
- - - 
1 noun phra~e analysis. This is now the second of the two 
pasallel phrases under csnshdesa tron, 
Step - 4. The phrase betore the postposition is analyzed by the normal 
nsm phrase analyeis 
Seep - 5. It is necessary to deterr ine what portion of the phrase 
befare the postposition relates exclusively to Noun-l To determine the left 
end of the Noun-1 phrase (e.g., in Flgure 3.8 below), we pick words one-by-one 
- 37 - 
from left to rtghk, and check whether each word can modify Noun-2. The first 
word found which eennot modify Wsm-2 is considered the left and of the 
first phrase (Noun-I. phrase). 
SHIKENUN -NO W -MO ENSAN Tt L' *'($tV 
taat tube in hvd~ochl~7ri~ acid c<-ppre" r 
(a) 
PLACE 
-/' 
thc hidrochluric acid in the test tuPc h\tdxocl~loric scid and cqp~~ 
and the- copper in ti test tube 
Figure 3:s T*'o different constructions accord?ng to the 
two different determinations uf the left end eP the conjoined phrase 
6. Words to the right af Soun-2 are checked to deteim&nr their 
- 
relation to the conjunctive firdse and its tonjunsts Checking pr~cceds from 
left to sight 
The. an~ltlsis of the fo1loi~';lng ph11~Le 1s Y11u~ltr;lt~d ~n Figtl~e 3,Q * 
RY lTEi4R7l1 -Ilu'O RJL7 -TL~ Fill' -YCl 1 -7 HI 
copper sulf id63 coppc~. (and) aultul EV~\; ,S tat ao 
the ratio hctween the smss of the cllrpcr .~nd thr sulfur ihich con.;titutc 
soppel sulf idc 
(See pgs. 39-40 for this figure ) 
3 4 Analysis of a Simple Sentence 
Japanese is a tvpical SOV language in whlch ACTOR, 
OBJECT and other 
case elements usually appear before the verb The construct ion of a t\ p ical 
Japanese sentence is shown'in Figure 3.10 (see pg 40) . A \erb ma\ gokern 
RYWWU -NO DOU -TO 
copper gulf Sde copper (conj unc tive pp---and) 
IOU -NO SHZTSURYQU -NO HI 
eulfur WBR ratio 
maning: tke ratio of the msees of copper and sulfur of copper sulfide 
(1) Step 1, Find the conjunctive yostpositiotlTU, 
succeeding part 
(2) 
Step Zc. Find Irom the succeeding part the nourl which belongs 
t 
to the saw category as coppez' In the abam 
phrase, the new 'sulfur' is found. 
copper sulfide 
I copper I 
mass ratio 
tempornralv dc eesglined 
acnpe of th eonj wne t i~ e 
plirase 
(3) Step 3 not applicsble 
(4) Step 4 Analyze the phrake bm the postpos ition TO 
copper = TO IOU -NOS '%f!I"L'SURYOU -EO HI 
sulfur mass ra i io 
ELEMENT 
I 
copper sulfide 
(5) Step 5, The second noun ~f tha conjunctiw phrase, 'sulfur' 
i cheeked against the leftmoat noun of the phrase 
krfnre the pastposition 'capper sulfide'. This 
nourn is related to the Eirwt now) of the ct~,snJ~artic"rIt~~ 
1 
phrasv, copper' . 'woppez su2f id@' f s alsa ~wtz 
to bc rr$at~d CL~ '1' T'ktis place8 tltc lest 
Zrntr~~dnrwv df the Wotrn- 3, pht ns~s 1 r3i~tetv tta elre 
1 
lrf t of capper ' , 
t 
Qb) Step b, me two nouns, copper1 and 'sulfur' , 1r-t the 
cbnjmctive phrase? are checked agalnst nouns in the 
? 
portion following Sam-2 Because the noun mass' 
can be related to onL\ ~ndividual ph\sicaI. objects, 
t 1 
the noun -3s' 1s duplicated f~s copper' and 'sulfur 
1 
The noun ratio' is relared t~ Q cmjunctive phrase 
as a tMyhole. Hence, we obtain the fo1lmiing result 
for the entire c~njo~ned phrase 
(mass 
UTR-AT I? 
copper 
d 
ELEMENT / ELEPEEIT / 
I- 
copper sulfide 
Figure 3.9 continued 
He (all eases) gas bums (INST) test tube (PC put in 
b-- relative clause 
liquid (OBf) heat 
waning: We heat@ the liquid which is qut it? the test tube, 
dodim chlorats (OBJ) pas hurner (INST) sodium cklosate (OBJ) 
heat teat tube CIW, PUCE, etc*) gut in 
waning: [Ssmonc) puts sodium &lorate in a test tuoe and heats it. 
(a) Ueuslly ehi~ phrase f s oairted, 
Figure 3.10 Typical Japanese, 
Sentences 
Fn 
1 
ra 
m 
Tsr 
'p3 
$3 
a 
f-t 
+a 
0 
m 
P- 
m 
PI 
r, 
3 
GJ 
m 
0 
G 
f> 
-r. d 
C3 
hi 
?-F 
a 
G.3 
"3 
G 
3 
'3 
9 
@I 
$1 
M 
a 
G 
iii 
6 
649 
3 
# 
ZF 
n 
cg 
t 
WI 
B 
pa 
0 
Ef 
From thrs table one can see that a gostposition in surface structure does 
not necessarily cdrrespond to a unique deep case. 
In the cuu~se of analysis 
wc muse assign appropriate case labels by considering the case frames of 
@ha win verb along with meanings of the head nouns of the noun phrases. 
h gostposition also plays the role of a delimiter which shows the 
sight b~undary of a noun phrase. 
The outline of the analysis of a simple 
~entenc e 18 FUI follows 
(1) At first the program look$ Ear a verb in ttlc input sentence, 
Because there may be embedded sentences which modify nouns in the 
win sentence, there 1s usually more than one verb in the input 
sentence. The program picks up the lef @most verb of the sentence 
(2) The strins before the verb is segmented by laeating pwst- 
posi t~ions 
(3) 
Slnce each segment is assumed to constitute a noun phrase, e 
is passed to the program which anaiyzes noun phrases. 
(4) When all ttac. segments are analyzed and tt~e hcsd nouns are 
determined, tt~e program checks each ~~oun phrase against. the verb 
nakltng wt~ehher a case rell=lrionship will be sat~sfied betweer1 the 
neun phrase and the verb The checking is carried out right to 
left starting with the phrase nearest the verb 
(5) When there are no more noun phrases to be ch~cked, or when a 
noun phrase which cannot be a case elemnt of the verb is found, the 
checking is tesmnated* If there remains an intrinsic case slot of 
the verb which has not been filled, we search for an alppropriate 
noun to fill the slot from the context This searching process will 
be explained in section 4, 
- 43 - 
We determine whether a noun phraw can be n case el nt Q vcrb 
by 'the following syntactic and semantic chuck 
(2) The cast) fraws of the verb. 
(3) 
The tw,ming of tho 'bond now uf the netm phrast. 
u.s$ng the second and t'hird hpre of infomution. ?YIP case &dot fillers in a 
case fx-~nw sf d verb are +reLatfVclv upper cunilept nouns, A sentence 1s 
considered to be an Phstantuti~n af a case Era*, and the nouns empla\ed 
will be lower concept nouns of the nouns in the cdse frames, 
Suppose we analvze the sentence* 
SHOKVEW -0 FITZV -NI TOUS Lr 
salt (OBJ) w'jter (IN, RFSLZT, TIHE, etc, wlE, dissolvt 
(Sowone) dissolves salt In w~ trr. 
We csn check tuthcthat the sentuncc w~tchcs the case S-IJW of f~XkiL! 
1 
The check~ng is carformed b~ considering whether salt' is a lower concept 
noun of 'material' , and whether 'water' 1s a lover concept noun of 'Irquld' 
Because a case fraw conta:ins only i1trinsi.c cases of a verb, we 
check extrinsic ones when a noun phrase is found not to be an intrinsic 
case element of the verb, That is, we check whether the postposition can 
mark the TIME or PUCE, 
and whether the noun phrase 1s an instance of the 
noun 'pldce' or ' time' . 
me above process Bay appear etraightfoward But sentences can 
have eeveral ~o~3rble interpretations for the fallowing seasons 
(1) 
A verb may hefe more then one usage (i.e., a verb may have 
several eaae Erams) 
(2) 
h po~stpositi&l can indicate more thgn one case. Sow post- 
pcab;itiona can oceus with almost any case; WA is an example, 
(3) Pk nwm wdified by an ededdad denterlee $8 ai8ualily o casu 
aloe filler of the e&edded abntem But we may have no 
syntactic clues as to what case to assign to the noun 
Xn the event of multiple interpretations the program derives labled 
inrerp~etati~ns showing all poaslble case relationships be tween specific 
now8 and wsbs, We cbeasse the intexpretati~n showing the preferable matchin 
of nouns and case by using an evaluation function below which has been 
establf ehed empirically. 
CFN : nuder sf intrinsic cases in n case frcam 
CIX . nuder st dntrfnsic case elemnts which are filled by the 
noun phrases in the sentence 
Ci! number of extrinsic case elements which are filled bv the 
ndm phrases in the sentence. 
C3 . number of intrinsic case elements which are filled by the 
noun phrases in the preceding sentences 
The value of this fmction indicates the degree of matching between a 
entesce and the case fzamg of the verb in question, The trial frame which 
gives the highest matched value is selected We then praeetad to the annlvr;is 
of the remining strings. If the selection is 
found to be wrong, duzing the 
succeeding analysis, control corns back to the point wt which rht* d~cist~7n 
was ~tade, di9eard.l; it, and C~QOSQR the pt~t tern #!rfrh giuw tt~t' t~c?~t highest 
~~~rtching value. 
XV CBNTEhTLbIL ANALIS IS 
4.1 Basilq r(pproac'lr to Centoxtual rhslv.iirr 
Our view of thQ ~TOE@SS of sentencc understanding is rouphl~ as 
follows. One reads sentences from left te right and understands then In 
succession. When he/she cannot understand a serrtence satlsEactorilv, he. she 
refers back to the preceding sentences ta rabtsln a ke, to understanding If 
he/she cannot find what is neeed , he/she leaves the questlon pending and 
proceeds te the next sentence If a phrase or n sentence 1s found wh~ch 
seems to solve the question, ttlen hershe checks wtr~ther ~t can renllr 
18solve the question. If SQ th~ PCII~E;'IICC is prwpexl\ organized into the 
previous contest and thc question is dismissed In anr caw the pending 
question is likelv to be disnrtssed as time passes 
We feel thls process of se~~tence ~tnderstandlng 1s not espec~allv 
complex. It can be reallzed through an ar~lrlclal lntelllgence approach 
Whlle we recognize that some klnds of problem may be solved only bv uslng 
complicated loglcal operatlons, we thlnk most problems In language under- 
standing can be solved by relatively s~mple operatlons Logical operatlons 
my only be effectively applled on o complete data base In which all the 
necessary axiom (corresponding to human hnotllrledge) are declared and no 
contradlctory asloms eslst, In the course of reading sentences, one has 
- i+6 - 
only partial knowledge about the context, and therefore, his knowledge 
is not coqplete. However, he can understand the meanings of sentences 
before he reads through the entire set. 
This means that one is content with 
incomplete deductions for under8 tanding sentences. 
For this reason, we 
employ rather than logicel operations, heuristically admissible operations 
whi~h tlBe an intarmedlate term mollory structure and various semantic 
tckartianehipka deeeribed in the dictionary. 
b ctsnc~ive of three type8 of memory, 
Long term mmisr\rr incorpordt~fi 
knowledge of the world, not considered here. Short term memory is for 
diate recall of unanely~ed strings under consideration, Inrcrmedistc 
term memory is limited but contains a structured representation sf recently 
analyzed strings and strings des analysis, We su rise our approach as follows: 
(1) Context i.; entered anto the ~ntermdiate term mem~fy, 
(2) 
Two k;lnds sf intermediate term memry are prepared. Cxne rs 
for representing the current contextual content, and the sther is 
to ~u~ttain pending quce;tions, '&he f~mer is furthe1 divided into 
the noun wtack (NS) and thc hypothet~eal noun stack (HNS) The 
latter 6a called the Trapping List (TL) . 
(3) Contextual. analysis is perf rmed after the processing of each 
sptactic unit such as a noun phrase or a sentence whlch conveys 
a mltary idea, 
(4) NS is organized such that thene words of sentences can be 
easily retrieved. Here ' theme words' mean the key sub~ects men- 
troned m the sentences, 
(5) 
Sometimes we have to refer to the succeeding sentences in order 
to understand a seotence. 
In such cases we do not iwdlately refer 
- 47 - 
to the succeeding sentences, but instead hold a pending question 
in TL to be resolved in the courqe aE analyzing th~ pr:,ccr.dit~p 
~eflt~ncc8 I 
4.2. kkm~n* Strurt~lra tor Cotztextuol lnt"ozr,ttatiun 
Tha analysis of B scntenrr is primrill p~otmd~d in the semantic 
descripticw.1 -- case Eram -- oS: n mit~ kcrh. Cr.rnteutual .tnul\si.= ib 111hzit111 
pt~~1ndt.d in acctlmulat~d inr o~mnntli~rt about t~olw~t.. 
T21~ .c7tsj~ct< ~t t~t~voptri 
that atr ttrr thsws .of thr scnton~~s, and what \~.ls bpt.n plcdicatcd QI thaw car 
usualkv be charoctexized in tcrms of the r.lQun,s appearing In the sentences, 
and these ~fFit importa~~t clues tot c~~~textual, rsnal\sis, 
tJe assign a daffesent LISP atom (produced b\ the LISP funct~on 
! 
gensvm') to each noun which appesrg Infamation about each is entered 
on the respective propestv list, The fldgs tabulated an Tsble 5, L are used, 
attached noun 
re lation 
I 
con tent 
LE X 1 Link to the dictic. 11 1~~1c.ll ~it>sc~iptloub 
SA'S R 
CASE 
PRE 
POST 
I link to thc case-frame In whlch the object appears 
llnh to any noun atom tah~ch appears In the prevlous 
sentence, and FJ~~C~I represents the same object as 
thls atom 
I 
the Inverse relatlon of PRE 
link to anv relatlve clause whlch modlfles thrs 
object 
'Irnk to any noun atoms trrhlc'n appear In a conjunct~ve 
phrase togethe1 wlth th~s ~bjsst 
- 
We can retrieve all the descriptions given for an object to which a noun has 
been assigned, We stack these LISP stom called Noun Atoms on NS and HKS 
- 48 - 
(A) Noun Stack (NS) 
When we start to analyze a sentence, we stack a list of noun atoms 
which are a~~igned to the nome in the sentence. 
These noun atoms are 
reordered ascarding to their degrees of importance, 
NS has the construction 
ehown in Figure 4.1 
( ( noun-atom-1, - - - - -p nom-atowi)(- - - - )(- - -1 
whi~h appeat in &he most rerent sentence 
Figure 4.1 Construction of WS 
TQ decide how importme a word is, we use the following heurlstlcs. 
(1) 
In Japanese a them word 1s often omitted or expressed by a 
pronoun rn succeeding sentences after ~t appears once. In other 
words, the bord which IS omtted 0s expressed by a pronoun 1s an 
important word for the understanding of a sentence. 
(2) A them word ~lav apyselar as "dubject1' in the surface stl~~cture. 
To emphasize a word whxeh is OBJ-case in deep case structure, or to 
de-emphasize a word in the ACT-case which is not worth mentioning, 
the pasgive voice may be used. This places a stressed word In the 
subject position of the sentence whlch would otherwise appear as 
object or xnd~rect object 
(3) 
The xmportance of a head noun in a noun phrase is greater than 
that of other nouns. 
A afqle example of ranking by importance 1s shown m Figure 4.2, zinc appears 
in all the sentences and is the theme word, 
Input sentence 
Nl W2 Y3 
WU"SSLB0 -MI 1OOp~Elcs S2.r21TSZtRYcJLf-8~J 
nrc.lting pot (PM\IE, TTbE, IORJ, atc,) WS~ 
NG N5 
AEM -0 IWTE CASV-8UNAd.a -DE 
zinc (@BY, TfJBJI put in 88% bun~cr (PLlcES, XE-\*$T,t?tz 'I 
NES SEi I, TOk!SWLTA, 
t~eat wlt [FAST TFNSE) 
hkming of the input sentence : 
Sl* (Swmewnc) put 10IOp of zinc An a mltlng ~ot 
S2 (Someone) heated it bv a gas burner, 
53. (Someone) melted rt. 
Changes of WS 
Repinn~np of the annl~$is of S1: ((Nk IT3 Eu'2 N1)) 
End of the a11~1vsI~ 51 ((U4 N1; W 3 K:)) 
Br@nninq nf the ,~nal~sit. 13t S2 ((h'51 (HA N1 N3 N?)) 
End of the ,mnL\sis of S:! ((W4 N5)(N4 Nl 33 E32)) 
Beginning of the anal~srs of S3 (NIL (N4 W5) (N4 Nl Y3 ~2)) 
End of the analvsis of S3 ((W4) (34 N5) (W4 K1 N3 K2)) 
Flgure 4,2 Changes of KS 
(B) Bypothetical Now Stack (HNS) 
We fir~t show examples which cannot be properly analyzed without HNS. 
(a) $mZ;6 -TO SMSO -0 2 -NO WARXAZ -BE KONWSHI, 
hydrogen oxygen (OBJ) two to one ratio in tedx 
KONO 
-WI ----- 
- 
KONWWLTAI 
this gas dxtu~-le (PUCE) - - - - - 
(Someone) intedxes hydrogen and oxygen in the ratio of two to one, 
CIp..II in this gas dxture - - - - - 
(b) SBOttl$N 5gr -0 UZU 100r;c -NI TOUSU, 
8alt five gram (fJsJ> water 100~1~ (IN) dis~olve, 
KONO 
- 
SUZYQmKX - Wh ----- 
the solution 
(Somane) dissolves 5 grams of salt in lOOcc of water. 
The soLaatir;kn is - - - - - 
In these two examples, though the deaons trative KONO (the, this) is used, 
the object referred to does not appear' explicitly in the precedinn sentence. 
The object referred to is produced as the result of the event which is 
oxpreared by rhe preceding sentence. As mntioned before, we append to case 
frama in eha wrb dicti~nav descriptions of any ~bjects which my be 
TQMSU (dieeolve) ha the ease frame: 
((ACT huaun) (0BJ material) (IN liguld) ) 
and this case fratne has the addxtisaal description: 
'aolutksn ('solvent (X LN)) 
('solute ( 0BJ)) )) 
The symbol $ in this description is a LISP function which fills the 
- 59, - 
sg,ecific ease elemnts indicated in the arg 
nt the currant malioat$a~ 
of the ease f ram, fie sentence 
dssociatad with the above caos fr msn~lts in aha FolSaqdiap f n tearlppset~t $,an 2 
a new object, a solution whose soavant is vstattl and whose golute is; salt 
msults, Wa repmsant this newly produced object in HWS instead u$ K$ 
for eha &ul%wtiting two msk;tlnrtl 
1 b the descfiption ia baaad un wcettabn knm*ledg~, fe is likely, bur 
not necessarily so that the object is produced in the redl world, If we 
find some descriptions sf this derived object in the succeeding sentences, 
we wiU decide it really ensts md transfer the reprasentstnon from HIS 
to NS, 
2, Because the newly pssdwced object is referred to in the swcceed$ng 
sentences sumtims by different wax ds or by sjmtar ti~:~l.v different 
form, it is soriieaient to stuck them individusllv in WgS, 
4-3 Estimation 0% the Omittad 'Gk~rds 
In the analysis of s Jspmesc sentence it is isp~rtant to supply 
omitted words drawing from preceding or succeeding sentences, TB do this 
we must be able to: 
1, recognize that a word is omitted and 
2. search for an appropriate word to fill the gag 
Our contention IS that an individual syntactic unit such as a noun phrase or 
a simple sentence conveys a definite idea; a noun phrase may designate a 
certain defirdte object, a concept, or whatever, and a simple sentence mv 
describe a defin teievent. In order that a simple sentence describe J 
definite even, each intrinsic case element of the ease frame must be 
- 52 - 
epscidicd by particular abjecte, 
We can detect an omitted ward by searching 
f~r unepecified case elomeete in 4 case frame. 
~oreoves, we can guess from 
tha ceiaiee Xsma what kind of nouns ahauid bo supplied tu fill. any gape, 
In thi~ mnnor wa can detect and supply omitted wards by using the 
@mantie d~ser%~~ia3es~ in tha dfctisnary. 
(1) bitead Wosds in a Singla Sentence 
men we have finished th~ %anaLyei@ of 8 sialplc zrrntence, FM chack 
w2rothar there raaata come intrfist~kc cages rw be specified. Tf there remain 
gme, we search Erar appropriate ftllsrs in the preceding aenteraceta, 
The 
~sarching psocess ia carried out in the forlaw lng way, 
(i) 
We aeasch through HNS first, because an object newly cseeted by 
the preccdqw event is sften the theme object of the present event. 
(ki) In Japeneee, idenrxcal case elements in succeeding sentences are 
apt ta be omitted. So the pteviow sentence is seasched Ear elements having 
the erne sea%le relotioar the oaa under consideration tks~tdgtl MS+ 
{Eii) 
If thc above psocassas fail, then we ehack the wzarda i, NS rnt al 
thg mrd~ that have appeared in the three previoua sentences one-bp-one until 
we Eind remntisally edmhsaible ward. 
(be 
If we cannot Eind a suitable wosd, we set up a problem in the 
trapping hist TL (mntioaed in the next section). 
Same results of the psa(ceesing are sh~m in Figure 4.3. (pgs, 54-56), 
(3) 
Oaitted Ward in a Bow Phrase 
A aam ts classiffed as either an entity word OX a relational mad. 
&st noma have definite maning by theraselves, and are regarded as entity 
words, Mowever, saw. kinds of now have relational meaning. That is so 
say, they have slats Za their maning to be filled fa by other words, in 
- 53 - 
(a) Input senteaca: 
diaee result of the analysis of the second sentence 
*final result abtained after searching 
Figure 6.3 
W IN -€I -NI 
naphthaline (OBJ, IOBJ) test tube (PUCE$ IOBJ, IN, etc,) 
I=* W+BMU -DEl MESSMITE, NUSMI, KA~SATSUSURU, 
put in gae burner (INST, NTHOD) hent mkt eabaerve 
waning: (Sowone) puea naphthalioe in the test tube. 
(Somane) hektta (it] by n gas burner. 
(Ssmana) mB te (the naph tkalf ne) , 
(9am~na) obae 8 (the naphthaline). 
*result ad the analysfa of the firsL atinknee 
(so'meone) nagh thaline 
8S=((N1 WZ)) HMS = MIL 
dfate nsuft of the malytiixs sf the second sentence 
gas burner 
Figure 4.3 continued 
**Shou& the third kod fourth seateanr als~ Ru 
mrkem, they are pr~wrhy filled in, Eel 
okr taiaed. 
Rgure 4.3 continued 
order that they exprese definite ideas. 
Sometimes a relational noun is used 
alone in a noun ahxase. In this case the relational noun must be semantically 
c~rmecttad with lother word8 which are o~tted in the prelisert~ noun phrase. 
Such 
cxamplea are ehown in Figure 4.4 balaw, 
(11) rov -Q NESSURU TDKI XRQ -GA HE~S~U , 
sulfur (OBf, TOBJ) heat when color (SUBJ) change 
meaning: When (eowone) keate sulfur, the color changes. 
The phrta~e '$RO -M is a noun phrase but it h8 incomplete 
collcar (s'tSf3J-l 
by itself. We can easily understand the color means 'the color of the 
eulfuat, 
(23 Efl$AH -0 -HI 
hydroehlorbc acid (OBJ) teat tube (PUCEb TIME, etc .) 
2Occ zmu. 
put in 
waning: (Sasde~n~) puts 20cc of hydrtachloric acid in a test rube, 
&The word 2Ucc i put in a separate position from ENSAH fiydsachloritl 
tac%d) in gha oentence, Zr, howewr, specifies an attribute of the acid, 
Vo;&m, 
4s the fdaab step in the analysis of a noun phrase, we check whether 
there remain relational nauos which have no definite meaglng. If found, we 
search through NS for words whrch are suitable to fill in the slots of the 
nome, The ecarchihg process 1s che same as for omitted words In siaple 
seatencea, Sometiaes the omitted words exist in succesding sentences, so 
we cczn set up a problem in TL, if we cannot find an appropriate word in the 
gracedhng aentencea 
(3) 
Detailed Description of the Trapping List (TL) 
lIost enaphoric exptessiom and olmittrd words are well onslye~d by 
searching throwgh the paceding sentences, 
Howav~r, ua need ametims to 
refer to succedine, sentences in order to anlrlyto a $anten$@ properly. Zhn 
sentences sk~aum in Figwrc 4 5 dre eesxnaples, 
manhng: - - - - the eoapound which is heated and whose state changes- 
(2) ONl3.I -0 Z"3"SEI -NI SHI, ATSLbRYOkC -8 
temperature eonstanf (PLACE, RESILT, ete 1 )OBJ) 
1 
increase when 
waning: When the t.i.mperiltwe is kept constsqt and the premure is 
bnereaaed, tkr volume of gas - - - 
Figure 4.5 Examples  here omitted vclrds appear in 
succeeding sentences 
Because the precedlng sentences have already been analyzed and both WHS and 
NS have been set up, i.t is easy to refer to the precedlng sentences On the 
other hand we cannot immediately refer to the succeeding sentences if this 'is 
called for. 
To solve this problem we set up a trapping list TL, The bas ir 
organization of TL is show in Figure 4.6, h trapping elemnt 'is a triplet 
I--.- -1 
e trapping elemnt 
ader 
FP t arbitran lfsp function 
F2 r erbitrary lisp function 
Figure 4.6 ConstructLon of TL 
and corresponds to s pending problem. 
When we cannot Eird an appropriate 
word in the preceding sentences for an omitted word or an anaphoric expression, 
WE put s new trapping element in TL. At this tdme the first of the triplet, 
N, is eet to zero, When a noun phrase dn a succeeding sentence is analyzed 
we pick up nouns isom the noun phrase one-by-one end check whether the 
present noun can resolve a pending problem in TL by evaluating the function 
Fl la the trapping element. 
We have defined several LISP functions far the functdoa Fl. These 
dunstions wark aa follows. 
(t) 
They check whether a new at hand Bsn salve the1 problems in TL, 
(ii) 
If it can do so; they update the dlta (for example, iE the 
function F1 is the function which searches thp words in TL for fdling in 
the? omitted case element, then the function will put the present noun in the 
case frame), and return the value 'DELETE' Then the system. will delete 
the trapping elemnt from TL. 
Ciii) 
If it cannot do so, the system adds 1 to N, the f icst element 
of tbe trapping element. Wen M exceeds five, the trapping element is deleted 
frorp TL. That is, it is decided that the problem corresponding to the trapping 
element can not be solved at ell. Before the deletion of a trapping elcmpenr 
its third element, the function F2, is eveluated. Thus far F2 has cnlJ b~an 
used to provide default values to sllowu some intet pscrsrlon 1 1s p~ndinp 
By using Pbp idea of TL, we can t;eparate *ridas checking wchuuisari 
fro& the win program. They can be invoked crutanwtitaliu whcn o noun &ppatn 
in n senten&@, The idea of TZ. msedles thee af 8. Z'harniak;'~: 'dsw4rtt (J197211, 
When his rryateffi sncowntars a coctoin word, far cxerirlc, 'pi~ bank', it cuoeta~ 
e demon which tries to catch from the succeeding sentences any word (e.g., 
money) related to the key word. Ke fear that unnecessary knowledge wilt 
clog the system with e 'combinatariel explosion' resulting from the proliter- 
ation of demons, Ous trapping elemat is gut in TL only temporarily to 
compensate for any missing elements to be retrieved from succeeding parcs, 
Hence the uli~tlcessary psoliferati~n af elements tgsv be avoided, 
4-4 Processing sf haphosic Expsessi,cns 
In Japanese anaphora is exprwsaed b\ using the articles KONO, KQRE, 
sb nt~icl~ correspond roughl~ tm ' the', ' this' and ' khese' in hpltsh. 
prflnoun KORB is used to designate o sinfile object in tha prereding 
sementes, and the pronoun KOUU is used to designate plural objects, 
The 
article E;OMO is used as a constituent of a noun phrase. Though the articles 
in English modify the Elrst succeeding noun, KBNO often modifies a no&? at 
some distance. h example is given in Figure 4.7. 
noun noun noun 
KONO SBIKm -NO NW -NO MU 
this test tube 1n copper 
the copper in this (inside of) the test tube 
this copper in the test tube 
Figure 4.7 
In thie exawple there ate three nouns following the article which 
can be wcfifled by it eyntactieally, We must decide the preferable modlfica- 
ti@ pattern by using c~ntextual informention, In the analysis af a noun 
phrase, we scan the words one-by-one from left to right, When we catch the 
article 'KDNLa, we put it fn the tewarasy stack. The wosd will then be 
checked to aee *ether it can modify a noun in the fallowing noun phsasa. 
Vhea.1 we scan rh~ now SH31aNUN (tset tube) in Fbgure 4,7, we check whet'taer 
the object indicated by it was elready mntioned in the preceding sentences. 
11 it was, then the article RON0 is regarded as modifying the noun ' test 
tube'. If not, the article is stacked again, In this way the article will 
btz checked against rne nouns in the noun phrase until the noun modified by 
The article KOWO is used in the folhawing two ways: 
{I) SANS0 -W ARU MONO SANS0 ---- 
- 
axygen (SW, ACT) exist oxygen (OBJ) 
mere ia% oxygen The ovgeaa - - - - - - 
%%a a~un SM$O mdfffad by the artdele HONO is the same entity noun which 
appears in the first sentewe. 
(2) %US0 -a UU, - KONO TAfSEKZ -0 
volusle (053.51 
There irs oxygen. The volume of the oxygen - - - - - - 
Zn thfs cae KONO alone desigaates the entity noun SAHSO wh:ich appears in 
the first sentence, This usage is permitted only if the noun modified is a 
selotioaal rr,wusls If the now has only a relational maning, 
the second usage 
appears =re of ten thaa the first, 
me 8aa9ashg deecsiptioas of articles and pronouns like KONO are 
prasotdully e~reased by LISP functions. 
The functions in the dicti~nary will 
- 61 - 
be evaluated if we find such words in a sentence, Tha fmctian lass RON0 
operates in the following way, 
(1) Mchech is mede to see if the sucq~rding n6un is icletlonrl. If the 
noun ha;% only a mlarftan&l warzing, kt 89 firsti sswumd tl;tat the artieka 
KONQ is cf the second usage and we go to step (3). Itf nat, wa p to 
gtep 3621, 
(2) Tha Elrs t usage of KONO has thc following three veoir'ties. 
(i) SMSO -GA ARU. KOHO WSO -0 - - - - 
There ia DV~P~. The ~xygstl - - - - 
The now mdifiad by the article is the same noun which appears in the 
preceding sentence. 
(ii) SANS0 - RRU. KQNO KITAX -0 
These is oxygen "IThegm---- 
The noun ' gas' mdifled bu the alticle is rn upper concept noun of the 
referait now 'oxygen', 
(iii) SANS0 -TQ SUISO -0 KONGOUSURU, RON6 KBNCBLKlTAS -& - 
oxygen and hdra&en (PBJ) mix gas dxtwre (OBJ) 
(S~mone) nixes oxygen and hvdtogrn, The gas mixture - - - 
The article modified a nodnalized form of the first sentence. The 
first sentence lgstantiates the case frame of the verb 'mx' FJe ewduate 
the NWS description of the case frame and obtain a new lnierenced object 
'mixture', whose elem~ents are the oxTgen and the hydrogen. 
The noun 
XONWUKITBI modified by the article is a lower concept noun oi the infesenced 
now (mixture) in ENS. 
According to these three varieties, we provide the following three 
check sautines. The order of checking in shown in Figure 4.8 
- 62 - 
Figure 4.8 The order of eheskbp 
(check 1) 
Ia there in the li~t the earn noun cre the noun modified by KONO.. 
(cherlr. 2) 18 t11ete in the liaat a lower concept noun of the noun modified 
by KONOu 
(check 3) Its there in the MMS list an upper concept noun of the mdified 
noun, ad are its patopert*s consistent with those of the 
wdified noun, 
Zf we cm Eind a naw which satisfie& one of these three conditions, we 
decide that it is the referent noun. If we cannot, the function for KONO 
returns the value NIL. 
(3) If the noun which fQlLaws the article has a relational meaning, the 
mkaninp descriptiwn of the now ha@ dore wh;d eh must be filled in by 
other wotde. What kind gf nsws is preferable far the slots ia described 
in the waning dee~riptiran~ We csearch ;tn MS and HNS for a object 
which satisfies the des'csiptian, For example suppoee the input is 
SmSO -GA daRU. KONO TAISWL - - - 
oxygen (ACT SUBJ) exia t volume 
The nswr TAXSKI is an attribute zlxolfnw So we look for a noun which may have 
the attribute and recognize that oxygen is appropriate. hother example is 
-- GA MU. KOMO -NI - - - 
test tube (ACT SUBJ) exist in (PUCE, RESULT) 
There is a test tube, In the(test tube)- - - 
1 
The ncrw WKA (in) is a pzeposdtiurtaS, noun which require8 B container1 or 
'liquid', We em easily racopite the test huleict se s Ssvar concept noun sf 
'clontainar, ' Therefore wc aasum tha word ROgO is uad f~r the test tube, 
Xf we find no such nouns, wet gupposc that the axtiel@ KOw i~ not of thle 
eacpbnd usage But of the first. So ua will ga tc step 2, 
The pronoun K03Ui (this, it) is used in sonteoccs ss e rasr cicrnt. 
case [fra~~ description of tha wrb in a aentetlcu. Tha pustpositivrl attechad 
to the prondun indicates a set of possible cases, By t&iog fro& the frams 
the cases which belong to the set; we can obtain the semantic descriptions 
which are satisfied by the object des~gnated by the pronoun So we search 
through HMS md NS for an ot3 ject which : atisfies the descsiptiens, Consider 
the following: 
kEZU 500cc -E$1 UU. - ROW -WJ, SMUKL%W 
water (ACT SLk3J)exist (PL!tF, WI%T, salt 
TINE - - - 
28% -0 ZERU 
(OBJ) put in 
There are 500ee sf water, In this (water)(someonr) puts in 2 grams 
of salt. 
The set of possible cases for the postposltlon NI is (PLACE, RESULT, TIHE, 
BEWFZCEHT - - -1, and the case frames of IRERU (put :in) have the cqse PLAw-, 
We can predict that the pronoun KORE (thls, it) fills the PLACE case in the 
sentence. The semantic description says 
that a lover concept naun of 
container' or 'Ilquid' is preferable as the PLACE case of the verb IRERU 
(put in) . 
The object 'water' , which is a lower concept. man of ' liquidv , 
is found in NS, and is 
detemuned to be the object designated bt the pronoun 
We have some other pronouns and articles in Japanese which are analyzed 
in ehe same way. We provide different LISP functions for different pronouns 
and put them in the dictionary definition8 of these words. 
T, Wfn~grad treated the saw problem in hie excellent system 
SkSRDLU (1971; 1972). lbwever, the world which his eys tern can deal with is 
very hjdted. Sn order to construct a system which can treat a wider range 
aE ecntcncee, the gystem should be equipped with the schema representing 
the re$~ei-srar%h$ps between event@ and abject (an event my iwly tl.u$ occuttrarentllc 
of new objects et changee in the propertie$ of sbjects). 
In real. world 
eeatencee, there exists mse complex ghenoena about anagheric expressions 
and ad~~ions of w~rd~ than those treated in SHRDLU. We do not claim that 
aut eyetern cm treat suck .c&pl.ex pnensmna, but we hope that our system can 
be evolved to cover such phenomena by man$ of combining contextual analysis 
procedure with seaantbc descsiptlons of words, 
kn the prevslrrus ~ections we described the sezwntfs and c~ntextual 
maJysPa pracedure af our @y@tQms 
In this seetLm we explicate by using 
sentences hw these fmctioaal wits are organized in order to 
analyze fai sky co~~~plex sentences. 
(I) Suppose the input sentence is 
&SHmPSAETE 'SAZSEKI -a MEW-SURU WKI -NO SAHSO -NO 
be eonpressed voLume (SUB5 ACT) change t I= OW gen 
when 
JOtk3Ca -0 SATSUSBI , SONO ATSWOKTJ -0 SOKWEXSHI , 
atate (OBJ) obsesye the, ire pressure (OBJ) measure 
SOW -8 G W -81 UU. 
it (0331 WGP~ (PLACE, RESILT) express 
CSoaane) obgervee the stare of the oxygen when it is compressed and the 
vdum (of it) changes, migsusesr the! pressure, and expresses it by a graph. 
The sentence is analyzed by the fo1law;ing steps. 
1 T'he program first tries tci find the Pefemst verb, am! anslyaes 
the rlausc governed by the verb. T%E sentence ASSHMFKWSWMTB (ba ruaprrsaedl 
is tknaly~ed first. Thia osntence llaa lan irm~uler ~tsuctwka in the RO~IBF 
that them are no aq;lticit wsa @laments beEow the '~"a?:rL kPB, vase olaw~nrs 
are omitted in thia aentan~.Fa: By ellacking ttto dnh$ectfuza QI the verb 
(ASSHUKU-SURU (to rampress) ---ASSHWUS83: (to be c~@presscdl). w re.e\.omici. 
that tho ecntencr is in the passive voicr. Tho lakical drsrrLptiaz~ of tile 
verb in the word dictionam indicates that it takes twc intrinsic rases, 
theat is ACT03 and QBJBCT. fn a Japanese sentence cspeefab&l; in thc field 
of chemistry, the case element ACTQR is apt to be neglected. Therefore we 
adopt a dummy filler fa the ACTOR to represent the author of the sentence 
or sow other human being. As there are no preceding sentences, we cannot 
frll in the OBJECT ease immediately. So we set up the pending problem in 
TL which will watch the analvsis of the succeeding a tsinps to fill the gap 
2 The clause TAXSSEKI-CA HENKA-SURW will be analyzed next The 
verb HEN&\-SURU (changed requires uln$v SLWJ case Thr! puseposfrl~lrl Gb, 
attached to the; noun TAISEKLCI (v.sLwm) paasiblb LiapLies the case SlrRJ. Ttlc 
noun TAISEKI: 1s a Irstver concept noun of 'attribute', which stjtisflcftzs the 
semantlc condition for the case element. So thls sentence is analyzed :in a 
straightfoswasd manner, However, because the noun TAISEKI 1s an attribute 
noun, we must find the corresponding entity noun. 
Tkqt IS, we must identifj 
the object whose volume is being referred to. As we cannot find such an 
object in the preceding sentences, we set up a pend~ng problem In TL, I3.1 
checking the inflection of the verb WENUSURU (change) and noting that it: is 
iwwdidt ely followed by a noun, it is recugnized that the sentence is an 
embedded sentence modifying khe following noun TOKI (time, when), Fu'e then 
- 66 - 
connect this caentential part wit.r the norm TOKE by using the relation SMOD 
(WDif ied by a Sentence) . 
3. Wen we analyze the next chaau~e, 
m3, -NO SMSO -NO JQWAJC -0 WSATSWSURW 
ti- QW$Qn $rate (OBJ) observe 
when 
ua first perfor. the analyais of the norm phrase TOKI-NO SANSD-NO JOVTAI. 
pemisaible becauee 'oxygdn' is e lowet conlcept noun of 'nrateriel' , and can 
be lodified by a word which designates a speciaii point of time. 
The noun 
mKI (ti=) ip mciified by the e~nteatid part analyzed at step 2, 
and 
deaf gnatee the tfm wbn the event expressed by the senteatid part occurs. 
The cedinotion af SWSO (owlgen) ad JOUTAI (state) is also permissible. 
The nouns T.OKI (time) , SmSB (oxygen) and JOUTAI (state) in the now 
phrase activate the trapping elements in TL. The noun SBLNSO (oxygen) 
aatissfiera the tloradftlcsns sf the two trapping elements set up by s tepa 1 and 2. 
That 94, 5M58 (ovvn) fills in the case OBJ af the first clause, TAISEKZ 
(vstum) in the wecsad clause is wgarded as the vdum QE the oxygen in the 
@u~tent c%&u~a. 
4, The next rlawe AmmYOKU-e;) SOKWEISHI presents no new preableus. 
Warever a referent for the noun ATSURYOKU (pressure) must be found. 'oxygen1 
kw the preceding sentence is easily fomd to satisfy the conditions for 
having the qualf ty ATSmYOKU (pressure). 
5. The remaining steps follow along sirmlar lines. me results of 
the parsing of the expressawn are shown Figure 5 1, (pg. 68). 
(2) 
The next example sh~ws how UWS is used. Suppose the input sentence is 
(1)- Input sentence: 
AS SEWUSMTE, 'lAISB#P -GA URZJI TORE -MO 
be eomprassed velum (ACT SUB.!) lchans *an 
SdWSO -NO JQmA1 -0 UTSBHIt, SOH"L~ kaBml9KV -0 
a xy gi$n ststre (QBJ) obaex~e the pna%pruz% (t7bJ5 
-wing: 
(Sawone) obserws the state of oxylpn which ia euqrc.sssli 
md whose volwmm changes, (Sumanel wasurns pte0;st3m 
and mgseaerrts it as n araph, 
Figure 5 3, 
SmSO -TO %MSO -0 KOHaWZ, RON8 KOMCQUKITAI -NI 
hydmpn md oxyen (OBJ) dx the gas mixture 
mu ~7x3 %MUBFSHIs KZU -GA DEKf RU, 
ftrs (if, when) explode water (SUBJ, ACT) be mede 
mrf dxcs hydtpgan md oxyen, and fires the gas mixture, then (it) 
e~lodar and watar rcltauhts, 
Ihr! Iollaving stapr are petforrd, 
1. Whan the analyofa at the Eiree clause SVISO-M SANSD-O KOM-EOUSiiI 
Ir caplcte, tha cue frrree of the verb KONWUSHI are ingtantieted. 
The 
S etpreeeioo of the epee fram which obtains the highest matched value 
iaf deterdncd. As the r$sult a new object ' dxture' is created and the 
elewnts of tbc plixture ere hydmgen and oxypn. This newly created object 
is put fnto HHS, 
2. The row phrase RONO KOSCOLKITAT-NZ (to the gas mixture) in 
the clause is mdified by the aaaphoric determiner KONO (this) which requires 
a referent. The now KONGOURlTbI (gas dxture) 18 a lower concept noun of 
'crinture' having aa components gaseaua objects. We search in the HNS and NS 
and find the object 'aixrurce' in HNS whase eleplents an the i~rdsagen and the 
3. Ihe object 'gas illixture' ia the theme of the succeeding senteaces. 
f t fills in the o~itted case ACT of the third clause a~d FROM ease of the 
fourth clause. Figure 5 -2 shows the result of the parsing (see pg. 70). 
Table b. I below showa the score ob taioed by applying our parsSng 
program to the sentences in a junior high school chemistry textbook, 
p) Input sentence: 
RON0 KONaWJTAI -HI 
this gaa~ dxtuw (QBJ, IQBJ, PUCE* ste. 8 &,mi ik7~m 
*Q11 
maning: IF (sna~ne) axes hydr~pn and oxyrn and Lmita. it, rh 
the mazaes uioh~ntl~ ad water hs pmdwe, 
TABLE 5.1 Successes and Failure$ Schema, 
V f WHlCLUS LON 
the our interpretive procedure as follows: 
[a) 
"%lrrou@ the use af gra tical case we describe patterns of activity 
in the wrb dittiranaq. The descriptions also contain information as 
ro haw activities are c~nmcted with each other and haw activxties change 
abjects. 
Eb) 
The: meaning clescriptioras or nouns am based. upon the upper and lower 
e~ncapt relationships and attrfbute value pairs. Some kinds of nouns 
are regarded as having relational meanings. Their meaning descriptions 
arc @idler to those of verbs, adjectives, and pt~positions. By using 
these deecriptions we can analyze fairly complex and long noun phrases 
where there ate few ayntactir clues, 
(I) We do nat use log%cal expsessrons to represent context Contextual 
informtiara is repreeented in the Eom of what we call intermediate 
tesm memsy. This in comlbinatlon with the semntlc descriptions of 
wards be3 errabled L~S to perform efflclent analyses dependent on con- 
textual information. 
($1 
We have developed a programming language which makes 3 t easy to ws:ite 
gw~~~xs for natural language ad to control the analysls procedure. 
- 71 - 
By wing this languhge, we cah i+neor;porate naturally sewntbc and 
contextual analyses into syntactic analysis, We do not aced a large 
&nd involved proernza which it9 rs~lponaible for the sawntic 
intamraeo- 
tian ~f tha output given by the syntaetle analysis eowonent, Im~tead, 
awlysera, 
We have obtained faf rly gaod re~t~lts wfth our npprasrtt', Thpia con- 
textual analvsds psggren on the other hand can treat unly local c~jntrxta, 
In ordet to treat mm gloBaS, contexts, @e lecl the foll~wing b~r~wm'~t%i 
(i) 
Fh must provide our system with an approgsfatesik~c~ c~rmspand- 
ing to human long tern memory in order to represent the state sf the world, 
'She system naus t haw fr.timst;wearks to express spatial relationship$ among objects 
time relationships awn8 evr.T7rs and so on. 
(ii) At the present stage we have only one .relationship CON t~ 
connect one activity with another. Wdwever, human knowledge of the world 
attlsnm~daeas varitaus, kinds of relsticansfzips among activities, such as 
cause, purposs;, mason, ete. These relationahips nmv play an S~purennt 
role not only in the analysis of sentences, but also in the ~nfercnce 
processes in answering a question. 
(lii) 
fie descriptions of verb meanlngs using case work rather well 
far analyzing verb-centered sentences. However, the results of analysls 
depend on what verbs are used in surface sentences Hence, the sentences 
which convey the same meanlng but are expressed by using different verbs may 
be transf omd into d-if ferent internal representations 
This is a serious 
drawback when cons tsucting ques tion-answering systems or other kinds of 
intelligence systems. In order to avoid this drawback, we attached a set 
- 72 - 
g rules to each case ftroe siprilar to the descriptions used 
at hsys;te@ (1913). We feel, howevesr, that thig wthod is 
rathsr rvku.rd urd char deeper structuraa ehould be employed (similar to 
' carc@ptU11 dependency' propoaad by R. C. Schank f 1973r; 1973h). 
(iv) 
In order that 1 ryeetrtn be able ka cumwlcata with paaph in a 
fLodbEc md nitural mntnsr, it 
t bs &la to drrim infemtrcas from 
incoqlece data baa... Thenrforu we mat: design a procedure othar than the 
mi fotr prcof pmcadurc &uch is the nraolutfon proof pracedum . 
(v) 
It flP nem~~aq SO apply out mtbed in fields different from 
chad~tsy and to reart *athat our oa tie dsecrigtion mthod should be 
&m@d or Bat. 
I"k-rsse are y ~chslars &as are interested in aing cme structures 
as a mpresentatlon of natural languap utterancee, 0, Bruce (1975) offers 
a good Gurney and a unified point of view in favor of case sys teas. We also 
Baliem that the erne system ia a pradaing apptawh ts the repmaentstion 
of amiogs ia natural imgurga. Fvrthpr va baliave that the idce of case 
giwar ua a weful tad ejot mptepsenting kwwledge of humn beings. 
"Sfre Ida= reported here cam from oly h;ources. We would especially like to 
tbmk It;, T~lxxaka; H. Tanabe and A. Tereada for discussions of cases and semantics 
Aha our thanks to D. Sdth for hi8 valuable suggestions; speeial.thanks to 
Linda Arthur for the preparation BR~ typing of the muscript, 
11 

References
Bruce, B. 1975, Case %y%tem far Natural tcsnguaga"', 9. iclf Artifdcdrnl IntsILigence, 

Celce, El. 1972. "Paradigms f~r Sontcnrr Recoplltlan." In SDC rlnal R~pol t 
~~T-l5092/7907* 

Charniok, 1. 1972. "laward a bdcl of Children's Stosv fc~rchrnalon". 
kMC-TR266, 

Fillworc, C. 1968. "T)IB Cnrie for Case". 
Sn hivrrs~lr: in Linpulstlr Tlleurv, 
(Di~ctr md Ham, sds. 1 , 

Nagao, b and J. Tsuj li. 1976. 
"PUTOH - P\ Hew Pmgraaning L#&npuapr lor 
Natural Language Analysis" , ( tosthcominp in AJCL) . 

Norman, D6 1973 "Newoori, howled* on4 the Answering nf Qoestiws". In 
Contemporary Issues in Csggitlve SswchsLogy: 
The Lsuoia Symposium 
(wins ton, ed, ) . 

Rumelhart, E. and D, No-, 1973. "~ctiwe Semantic Metworks as a lbdel sf 
Huwn bkmsy" . Psoc * IJCAI-73. 

Schank, R. 1973. "fdentifl,cetion of Concep tuolizat ions Underlying h'atural 
Language". In Computer bbdels of Thought and Lonpu~ge (Schank and folbt , 
eds.). 

Schank, R. 1973b "Plarpie blewry, rbal~ sis , Lsponse . Generation and 
Inference om1 English", Bruc IJCd!%-;3. 

Simmons, R, 1973. Seauntic Netwurks. Theit Computation and t'sc fur Vndcr- 
standing English Sentences". In ramputer Sbdels of Thought JII~ L~IIYUBPB 
(Schank and Colbv, edsd 

Simmons, R, 1975. t he Clown ELicrovorldr' . In Theoretical Issues in Satusal 
Language Processing (Schank and Webber. eds 

Wilks, Y. 1975 
"An Intelligent Analyzer and Unders tander for ~nglish" 
Corn. ACM 

Winograd, T. 1971 
"Procedures as a Repre~entation for Data in a Computer 
Program for Unders tandlng Natuid ~anguage' 
EWC TR-84. 

Winograd, T. . 1972 Under~tandinp. astural Language Academic Press. 

Woods, W. 1970. Transition Network Grdanoars for h'atural Language h\nal\sis 
Corn, ACM. 
