Towards Machine \]ranslal; on U. inf  C, onC(,×t, m,1 hfforma,tion 
Tim Cornish, Kimikazu \]"ujita, \[lyochi ,~uginm,'~ 
{ ti,n,fujil, a,sugimura }((fisl.mei.co.jl) 
M;~tsushit, a. l';lectric Industrial Co I,td 
Information ~md Commu' nmaLl()~' ' is Technology lmbora.tory 
Abstract 
A proposal is made for the use of contextual in- 
formation in the machine translation of Japanese 
and English. This paper describes the use of a 
Context Monitor to nlaintaill contextual infortmt- 
tion dyn,'unically and the ~tugmenlalion of appro- 
priate features to ~t semantic network to enahh~ 
simple inference. The al}proach taken is that {}t" 
"best gucs~?' processing with the cont~extual infor- 
mation being hal~dled with semantic inf{~rmalion 
on ~ shallow level. 
2 Introduction 
Current Machine qh'anslatiou (MT) systems proc,~ss 
input sentence by sentence. I\[owever, experience wil.h 
English and Japanese has shown that some languages 
difl'er to such a degree that sentential translation 
yiehls poor results, l,eL us first compare the results 
of a conventional MT sysl.em with those we expect, t,o 
get for MT with context: 
t. J::q'lJf\]~-._-i~¢{:¢/,?)\[ b v,-)-- 1/~3-'.~E~gfi £" 2)i L v, K 
2. K-7: ~t- laJaS'd?,l~ ~a ~t~.I;~y b t:: o 
4. k-c g \];! ( 5'~ko 
This might be translated by a current, machine tl'ans- 
lation system as shown in Figure 11: 
It can clearly l)e seen that meaning in IHally seI/- 
tenees is obscured. Let us compare this with I.he re- 
suits of a system using simple cont.exl.ual informal,ion 
ms shown in Figure 2: 
This secol/d translation is i-tlllch Ill(H'(? CO\]lO\['{~ll{. ;IH{I 
better preserves the meaning of the original se,lten{'o. 
An attempt has therefore I}een made to solve some 
of tile problems of translal.ing languages SllCIt sis 
Japanese and English using contextual information. 
Due to \[.he consideral.ions of wanting to produce a 
high quality small-sized MT system, lhe approach 
taken is to use tile resources awdlahle in an exisl;ing 
MT system and to process the contextu;d i,l\['orlmd.ion 
l There is obviously n great difference in results Imtweet, sys- 
tems, hnl, l.hese translat.icms relweSent tyl}iCal {uHe, llted) r~.stdts 
fi'om a numher of systems, a) and I}) options,hq}end on the 
default settings of individual systems 
1. The Chief I)evel{}pment li;ngineer develot}ed 
two new TV models and Four new video ,nod- 
els last year. 
2. a) A vi{h'o was shil)ped to the Sales Section. 
b) We/Soineone shil)pcd a video to the Sales See- 
l, ion. 
"I. a) 'l'wo models were released straight away. 
l}) \\/e/~olneone released I,wo models straight 
a*,v a.g. 
d. It sohl very well. 
Figure 1: (jonventional M'P Results 
I. The Chief l)evelolmmnt Engineer developed 
DA'O IICW 'FV lllodels ~llld \[()sir flew video n/ori- 
els lasl year. 
2. lie shipped the videos l.o t.he Sales Section. 
3. They released two models straight away. 
4. Tlmy sold very well. 
Figure. 2: Coutext, ual M'\[' Kesults 
on a shallow level only, using the information gained 
to guide the translation on a "best guess" basis. This 
kiml of feal,ure with rat.her light; processing for the 
production of a higher quality translation is desirable 
in a pracl,ical MT system because the advanl.ages of 
large-scah~ processing for deep conl.extual ilfform;d.ioi, 
are likely to he limited in Lids apl~lical.ion. 
3 The MT System 
The transhd, km system present,ed here is ;t model sys- 
tent whidl is being used to investigai, e t, he techniques 
proposed. The translation Imr~ is carried out h, Pl{.O- 
1,0(\] usinp; nn l,FC-LIike gramma.tieal formalism 2. 
'\]'he current, dictionaries conl,~dn information to trans- 
lal.e aboul; 300 words. There are 350 grammar rules 
which (:own' a wide range of sentence pal,terns. 
The context monil,or operates using information re- 
trieved from file f-sLrueturc of ;t SOlli.eltCe al't(W analy- 
sis. This ilfforlnal.ion is then used during the transfer 
2'\['\[le cn'igimd lm,gl.anl for l;;hgllnh-Spanlsh iranslallcm de- 
veh,p,,d I,y (1, Amm'es \[:\m, wes'89\] has been widely adapted 
and eldal'ged Io .I-I'; & E-J Iranslallon. 
57 
of tile source f~structure to the target f-structure. As 
context processing is carried out on only a shallow 
level, only information for lexical item, nurnber, per- 
son, gender, case role etc is used in the context system, 
along with semantic inIi)rtnation from the semantic 
network. The way that this information is used will 
be explained below in regard to the specific problems 
that tile use of context is intended to resolw~. 
4 The Context Monitor 
The context monitor proposed in this pal)er 
uses a standard focussing theory as a basis 
(\[Sidner 81\]),\[Sid,ler 8(i\]), amlough sonrexvhat simpli- 
fied according to the bes* :press approach that we are 
adopting. It is planned to increase the COml~lexily of 
this initial algorithm to reflect more current wn'sions 
of the theory as the system is developed. 
The context monitor has a number of basic data 
structures: Current l;k)cus, Actor l'~oeus, Potential l,'o- 
cus List, Potential Actor I;'ocus List, Discourse Seg- 
ment Stack and Actor I:ocus Stack. There is also a 
Current State List that maintains a record of all the 
semantic items currently hehl in any of the other data 
structures and the semantic features to whMl they 
are linked. This list is updated (entries added :rod 
removed) after every sentence. 
In ordc'r to lilnlt the scope of the context informa- 
tion required in the context monitor, an analysis was 
made of the main differences between 3apanese aml 
English that provide problenas for MT syst.ems. The 
basis of the analysis was to find what information can 
be gained from context to solve these problems. 
4.1 Plural Fomns 
Japanese is (in general a) mmlarkc'd for mlmber. I';n- 
glish, however, differentiates between singular and 
plural, q'his fact causes problems when translating 
t¥om Japanese to English as the mlmber information 
required for the inflection and declension of English is 
not available from the analysis of the 3apanese. I:or 
example: 
(I,oy(s),,,~, dog(s)o,,s lille) 
The boy likes th.e do~j. 
The bo W like the cloy. 
The boy likes Zhe dolls. 
The bow like Hie do!Is. 
Bogs like doas. 
In the current system an initial sellt.ence analyzed 
by the system is processed to find possible: ff)ci. II,ents 
which are in the l)lural or are in conjunction are stored 
as a set. 'Che set as a whole is given phlral nultll)er, 
aNote the use of .t'~ ('l(tchi') with mainly people and anlm.ls, 
but consists of imlividual items or, as} in the case of 
"/we ~e,v 7'V models and fo~tr 7~ew video models", as 
subsets. Subsets or individual items within the sets 
are available as antecedents to subsequ.etlt referring 
expressiolls. 
'\]_'hus, in the example text. in Section 1, after the 
initial sentence is analysed, the proposed focus is two 
new TV models and fmtr new video models, the strut.- 
lure o\[" which is shown in Figure 3 below: 4 
\[sell : \[set2: 
\[pred :terebi (T V), 
lllllll~l)\[lll' ~ 
nlod:\[p red:l.:ishu (model), 
spee:\[pred:,li (Iwo), 
_.\]1\], 
ref:set2\], 
\[set3: 
\[pred:l)ideo(video), 
lllllll;phlr 1 
mod:\[pred:kishu(model), 
SlmC :\[pred : yo,,(fiJwl 9, -.\]\]\], 
rcf:set3\], 
,'of:set 1\] 
l:igure 3: PIi.Of, OG Structure 
Sentence 2 (S?) is analysed and a test is made to 
see if any items in that sentence confirm or reject the 
proposed focus. 'l'he structure for the item g'7"~ 
("a video/videos") is matched by mfiIication with the 
structure for the prol)osed focus and can Im matched 
with a subset of it., namely ~)i k. v,/2"-)~ ~ - IJtl,~K~{\[( ("d 
new video models"). That item is there.fore taken Io 
conllrm the proposed R)t:us. 
That proposed I'oeus is, however, hnmedhttely 
PUSlled onto the focus stack because the subset of 4 
vhleos is taken as the current \['ocus 5. 'Phe item 13-)~";t 
of $2 inherits the feal.tlrcs of the set. of videos fronl S I. 
aml is therefore expressed ill the I(nglish with a plu- 
ral form: "videos". It is hoped that in this way the 
conl ext monitor will be able to distinguish between 
singular alld plural in at least some cases. 
II/ sonic cases t, here is Iio way of disl.inlguishing be- 
tween singular and l/lural rei~renee in .Japanese as in 
tile case o\[" the sel/tellce below: 
"/'eve (~)1¢1 llm~al,:o bO,:lht a ca\]w.. 7'l.~ey 
ale it "i~ lhe park. 
Taro aT*d lf m*ako I)o,:lht (some) 
('ak(~s. The?/ ate them il~ the park. 
'tln a shnplilied fi-wm, showing relevant delal\] only. Italics 
&l'e tl'&llSl&liOIIS for exl)lanation only and do not, appear in the 
s\[ t'llC\[ u l'e |)l'Ol)el' 
5q'he syslem currmllly deals only with local foe.us - there is 
no account (if gl.ba\] focus 
52 
Ill stleh cases L\]le e()llLexID lllOIiil,of Cal\]liO~, resolve 
slnguhlr or t;lural and so the MT system del'a/llt> will 
be relied on. l\[owever, l.ll~ conLexL nlo~lit,or at. leasL 
allows 15r coherence with Sll\])Se(\]tlelllL, \]srollOllllS. 
4.2 Translation of Pronouns 
.Jl_tplineso make,q llltich /1se oF l.he zero I)l'OllOlll/ 
(marked hero by "O"), especially in 1.1w sulLiecl, po- 
sil;ion, bu~ equa\]ly for other roles. I.'or exalIiple: 
(~ ¢ lIIOIl tabeLc shillla\[l.a) 
(O~,d,j Oot, j ah'eady eel A UX\] \[ 
U) haw al,'e,,,O e<a,,,, (i 0 
(:rh~,:D h<,~,~, ,~,'~,,,,0 ~<a,,,, (~/,,:,,,) 
(H<) h<,, ,~,'<<,<0 e<a~,, (~h,<,,,) 
This illelitllS that. i,hore is no illfornlal.ioli availahh, frolll 
l, he single, sellLellee Lo aid the choice c.\[ eq/iiva\]0iil. I'\]ii 
glish \]H'OllOllll (which i\]lllSl, iI()l'lli;ll\]}, l;l! I!xprl'ssed). 
As showll ill 2 and \[I o\[' t, lm exalnl)lC Lexl, in I"ig;m'e 1, 
M'F systelns use a, nullllmr o\[ niol.hods I()add all ovi,rt. 
l)roliOllii> of ),on involviug I, hc' riser in l.he Iili;tl choice. 
IL is claimed t>hat, if Lherc is a l)roliOtlli in ~ Selitellce> 
it; ll\]llS~ re\[er Lo tile foells of that l, exL se~lllelll, (ill or(lor 
1,o continue, the current segment) and it' there are more 
IL\]I~II/ ()lie prOllOllllS, ~l{, letlsID one ()f t.helll IlltlSi, re\[or Io 
t, he focus. By tol.3Ckillg l,h¢~ fOellS of H t,ext~ Seglllel/l,, 
0 |srollOllllS ill .\]~'/\])illlese, shollld he able (,(; he rescflved 
so LhaL il~li approl)riaLe overl, pl'oliOilli in ll;ll/.,;\]i~;ll ceil 
be selected for Lhe Lranslal, ion. 
'~V\]ieii a zero pl'Ol/OUIi iS deLect.ed iii a Seli(,l!ll(:c, il'ali 
aliLecedent, Call \])o foun(\] for ii., al/d L\]lal, ;liil.ect!(\]ellL 
is a, seL of it.eHis~ the= ov(~rt, l)roiiouu inslu'tc(I ill Ihe 
\];',ngllsh will I)e plural. 
Thus in '1 of/.he exainple from Figure 2, we soc thai, 
l.he zero pi'OliOtlll in " 0~,d,j <~: ~(~) j{t ~ \[/l~TL/.: " is alia\].. 
ysed as ref0rring Lo Lhe D.vo video models released au(\] 
is L\]ierefore Ll'~tlISIaLed wiLh ;t plur;d \])r(lll()till: "7'\]tC:\] 
sohl verT! wel?'. 
Noi, e, howew;r, thai. t.here is allil)i<,,;uily hi ,%'iilellce 
3 t)el.wcen whei, her l.he Zei'o I)rc, llOUli ill O \[.\[~{~.~/l(~<~; J 
C ~'.'~5'd L./.: (0 released t~vo models .,sl,'aigDI a,,ay) 
r0fers t,o Lhe Chief l)ew~'\]opinenf.. FAlgiIicer or l}ie Sales 
~ecLion. 
\'Vheli faeed wil.h anl\])iguil.y SllC\]l i/s l.\]iis> l;u'g<!~sc;i\]e 
a, LLenipLs aL coid.ext, undersl.anding lnig, hl. use in(t!)'- 
ellCC plans LO solve Lhe aill\];igtlil.y. \[Iowever, 1)OC,:tlISO 
of )>he lhniLat.ions of a Slii:,l\]l size MT sy~;lelu allcl i.h(~ 
feel Lhal, eveil \[al't~e s(;ale deep level seniaul, ic l)ro('ess- 
ing has noL heeli saLisfaetorily realised lksr linlilniLed 
donlains (wit,h which our lkl'l' syst.eui is in)elided 1,() 
work), we decided t.o atteilllst liuiit~d infereileillg hy 
the addil.ion of sortie features aim links I.(5 i.he selllmi-. 
Lic network oft.he MT sysl.eln. The hff~,'encing ahle 
Lo be lmrfornio(\] I)y such a niei.ho(l is quite simple> \])til. 
is hoped 1.o I)e suflicienl, for our nee@ in accordaliee 
wiLh l, he besl guess policy. 
4.3 Selilani.ic Networks 
~emantic networks are has)tally a hierarchy of con.. 
cept.s which are linked t,o one alloi, her in a neLx(,orl< 
t, ype sLrllcL/ll'e. ~elH311|,ic Ilet, works were ilfl,rochlced 
by Quilliau in 19\[i8 \[QuilliaN (ill\] am1 were wldely us~d 
in aLl.empt.s aL l(uowledge Ilasod ~ys\[elllS, imrticu - 
larly dm'iltg the l;ll.e 1980s. 
As all ex:uuple o(' suc\]l a system leL us hriei<ly 
consider I.he syM.elll I'or Japanese-l".nglish tl'ausla- 
I.ion using couLexl.ual iuh:~rnulLion prol~osed hy I1. 
isahara aml S. Ishizaki (\[lsaha,.a 86\], \[lsahara 87\], 
\[lshizaki 89\] and \[\[sahara 90\]) as oue l(nowledge 
llased apl~roach aml compare it, to the Lechniques used 
in i, he system proposed in I, he curreni, paper. 
The l,ranslaLion sysi,eln CON'I'll, AST tr;mshlt,ed 
,lal~;mese m~wspaper ari,icles into EnglMl. llowew~r, 
a major di/l',,rence r~gardiug our syst~ln is I, hat con. 
lext, umh'rsl.al.ling illvolv~,d analysing a sent,ence A 
(,'r; with au (~vort, ~.\]h\],,ct) and a s,~m,em',, I~ (,~g wilh 
a ~:ov,.rl sul>.i,<'t) ;rod thel, mal,&i.g I, hese SellHmc~s 
againM, ;~ immbcr o\[ Selll,l!llCe \])al,l,el'llS. \[\[" i~ llla\[,c\]l 
was I'omid Lllese would form a l,(~xl, paLLern C, wit.h A 
aud B as sulfl)arl.s. The snhject of A would he used 
(if slfil,able) 1o provide the subject lbr IL By adding 
\[lll'Lller Sl!lll,ellce.q alld t.ex\[ I)aI.LQI'IIS, it I'ei)resellLatiOll 
of Ihe entire Lexl. woilhl 1)e \['Ol'liied all(t I.hi.~ l.exl, rel)re 
slml;/l.ion iransial.ed iilt.o ;ill \]';uglish equiv;llenl, 1.exl.. 
IIow~!vei', I.his Coclnlique relies eli Lhe f;icL I, haL yell 
Call predicl. ;ill I.he i.ypes of Selll.ellC(~ t.hal, will OCciir 
;lii(i how I.hey coli~l)inc I() form all ellLii'l! text. (per 
h;ips I)~JssilJle li>r Lh<! l.ype~; of newslmpor ;u't.icles l.ii0 
CONTll,AST sy.';Leili ailile(I t.o l.lUilisial, e). \]lowever, if 
tl SOlll.011(:e Cllllll()l, fit inLo olle of l he l)reprepared pal. 
I\[ei'liS, ),he syslelll will t?dl. Oiir sysl.elil is inLended for 
lllore gell01.;l\] l/lllgjlla<t./,/tilt\] as We CallllOl. predicL Lhe 
lenglh of a I.exl: or whal. kinds of selil.0lico will occur 
v.'illliu I}lat. (exL, (.he (;oniexl. Monil.or lsrovi(Ies o11.. 
<u;oing; colilcxl.ua\] analysis wiihouL F, rel)resllliiill<t.~ tim 
lellgLh or nat.ure of l.he l.e.xl,. 
('.ON'I'IIA.<gT ills(; i'elies eli lilakilig\] ;I represelll.;i. 
t.ioli of Ihe elil.iro I.cxl.. Ill Ollr sysl.Olli I.h(~l'e is iio Illl- 
defsl.;lll+,{ill/~ O\[ I.\]ll) over;ill l.exl. M.rllCIAire (accordili,t.; 
it) oiir shallow hwel alllSi'oach ). liisl.ead> l.\]ie oh)eels 
alld evelllS referred 1.o ill t.\]m i,exl, al'(~ ~/liaiysed and 
IIl~lde avail;ihle t.o i'esolw, ~ SllllSeqllelll, analysis prol)- 
\[elliS. The t.raliSi;iLion i'OlHaiilS Sellt.eilce \])y SOllt.CliCC~ 
all.hough tile ~,elier;ll coliLexl, of Lhe Lext. is lllOllil.ored. 
Fiually, ~oinaiit.ic. Nei,worles, sucli as I.hal. l)rOl)osed 
by lsahiir;i eL el., )ire sl, al.ie n()Lworks. The links do 
lioi, chalice I)cq',veen nodes. The possibh~ pal, hs Lhal. 
are av;iihll)le l.hrough I.he nel.work lilay (!hal)g0 hut Lile 
links l.helilselves do ilOL challge, hi olir syst.el~u, t.he 
Imsic SC'IllalII.i(; IleD.v()rl¢ is si>ai, ie, defining irreflll.al;Ie 
relal, ious Ill!Lweeu I.he cont;epl.s in l.he hio!.'arehy, bul. 
eli I.o I) of this, ol.h0r linl~s are augjllieliLed Olil.o Lhe 
~J 
network and these links can change dynamically in 
respect to the specific objects and colicepts referred 
to in thc text. This provides a powerflil augnlentation 
to the basic nctwork. 
4,4 The Augmented Semantic Net- 
work 
Tile semantic network in this system is basically a hi- 
erarchy of Objects, States and Events. The addition 
of features to the semantic network in effect adds links 
to the network. Two kinds of link are proposed: per- 
manent links and temporary links. Permanent' links 
are conditions that must be true for a certain action 
or state-of-affairs to bold. The other, temporary, links 
are used to create a default state for the objects men- 
tioned in the text. As the text is processed, these links 
may change, so that the information awtilable to the 
system will differ from one sentence to the next. 
4.4.1 The Links 
The division between Objects, States and Events is re- 
flected in tlle type of feature given to semantic itclnS. 
For exainple, Events typically coilgain features about 
the sort of things that are affected by that event; 
States contain information the types of objects that 
may be in that state; Objects contain iilforlnaiiou 
about any subparts or if they themselves are typi- 
cally part of another (larger) object and wliat type of 
Event they are typically involved in. 
On this basis, the following types of link are pro- 
posed: 
. Condition (=c): (permanent) a condition that must 
|lold for a State or Event to come about. 
, Before Condition (BC::): (pernianeut) a condition 
that must be trlie before an Event or State conies 
about. 
* After Condition (AC::): (permanent) a co,ldition 
that becomes trtie after an Event or State comes 
about. 
* Has Subpart (l,as): (tenlporary) an Object lias re-- 
lated subparts or is a subpart of another Object. 
* Characteristic (has Semautic.Label): (I.emliorary) 
an Object has the characteristic of Senmntie_Lal)el 
(usually an Abstract_Relation: Size, Shape Colour 
etc.). This takes the form of: "lteni has Senlau- 
tie_Label", such as "Pelerhas Existel,ee Lif~span". 
Tiffs states that an item witll tile semantic item Pe- 
ter has an existence of some kind and filrther locates 
tllat itenl on a path of the network to the abstract 
relation of lAfespan. Iu this way, nodes between these 
two points are all available for reference by the infer- 
ence system. 
. Ability (able_to): (temporary) This is not fully de- 
fined in tim current systenl but represents character- 
istic features of items e.g. "dooF' often appears in the 
tlleme position of tile Events Open aud ,O'hll\[. 
These liuks are considered sutllcient for tile current 
capabilities of the system. I,inks may be deleted or 
others added as the range of the sytem widens, if this 
is tllought necessary or desirable. 
4.4.2 Permanent .(k Tenlporary Links 
The dlfference between perrnallent and I;enll)orary 
links is iu the nature of the iuformation that they con- 
vey. Permauent links are those that are augmented 
to tile lletwork and COlllleCt uodes cue I,o tile other 
in accordance with the features found in those nodes. 
~Bef(~re Condition and Aries" Condition links are per- 
manent, although the hiformation contaiued in the 
uodes that, t.hey connect to will only become awdl- 
able to the context nlonitor iu accordance with the 
tense al/d aspect of the verb (i.e.. all After Condi- 
lion is obviously only valid crier the completion of 
the (for exalnplc) action denoted by a verb has fin- 
ished. Teniporary liuks are those that supply default 
illforfllation to the COlltext Il~Ollitor coucerning nodes 
that it is COliCerlied wii.h. 'Fhus, for exaluF, le > an entry 
for a bird iuight state that it is Able_To Fly. llowever, 
if the input text were to state that a particular bird 
is unable to fly, that A1Ao "l'o link would be cancelled. 
Thus tolllporary links provide the iliforulatioil that 
the colitext nlonitor rises> using the teinporary links 
to spread throughoilt the network (within set search 
constraints) aild gat.hering inPoruiatioii that Call lie 
used for infcreilcing. 
4.4.3 Exanlph; of the Augmented Foat;uro.s 
All exalllple Of the features used to a/l~_~Illent tile se- 
niantic network can be <tivoli tlsiilg the e×alrll)le: 
Peler heard that ,\]olui had died. lie was very 
::3 a d, 
(-liven the dict, ionary entry shown (here sirnplified) 
I)elow, "l'clcF' will lie analysed as a male proper 
11OI111, 
dic(n ,'l>eter',\[selu f'eat:\[hmnali:yes\], 
i)n'opcr:yes, 
L~t~\]) d (!r :lu as(', I 
l)r"d:peter\]). 
When the enibedded clause is analysed, "Jolin" will 
be analysed in a similar way. The semantic feature 
human:yes locates these l, wo lexical items as sub- 
surned by the semantic feature "Living" in the net- 
work. Augmerited Featin'es Ibr a mah; bureau such as 
the objects relZ'rred to by the nanles Pcler aud Jolin 
are shown below in Figure d below along with possible 
entries for the l,\]w!nt die and the St~te be sad, 
It ca,l be seen that one of the Before Conditions 
o\[' the I'\]wmt Die is I,hat the actor role is filled by 
all il, em I;hat has the seilialllic feature "Living". The 
del'ault asstllUl'd, iou for "JoDll" is thai. lie is I\[llman 
,54 
1F 1 F John; Peler lie Sad h~tt Rlml~ Anthropoid ,~Klmri~sccr =, Altlmate \[Imt l~ittcnc¢ l,lfespan~\[ Die IIC:: ~cto," lAvllql tiC:: IICII~ Tr~lll~lellt'e 
AC:: actor not l,lvlng 
k(?;: actor (?orp~e 
Figure 4: Augmented l"eatures 
and therefore l,iving, llowever, the After (\]ontliLiol;s 
of the Event Die cancel the feature laving lil connec- 
tion with "John" ('~o1' means that a \[?ature aml all 
tile other features underneath it in the tree should 
not be reachable by that item), and state, that the 
item should be associated with the fimture "Corpse" 
(a semantic label in the system for something that 
was living but is no longer). Thus the semantic item 
"John" is first linked with the semantic feature "llu- 
man" and all the other featm'es inherited from that 
feature. However, the features associated with the 
semantic item "Die" cause the links associated with 
"John" to change. This means that when tile see- 
end sentence is analysed, the possilfle candidates f()r 
the exl)erieneer role of the semantic item "lhLSa(f' 
are analysed, an itmn with the semantic feature "All- 
inmte" will be sought, and so the item ",/ok." will 
not be considered in the. search as it is no long,-r on 
a path reachable hy "Animate.". "Pete/' is therefore 
the only possible antecedent. 
4.~ Articles 
Japanese does not use de.finite and indefinite arti- 
des and so when there is no ow.'rt determiner in the 
aal)anese, one must be supplied for the English trans- 
lation. For example, Sentence 2 of ore" example text: 
155" M" 1:111~3E l~ll V- ~l~',Ij: U/:: ,, 
They passed the videos to the Sales ,qeetion. 
Where a simple default rule is used for articles, ibis 
eouhl equally be machine translated as: I\]te 9 llaSsed 
videos to the Sales Division ~, where it can be co.sid- 
ered that some of the sense of the original seuteuce is 
lost. 
While the use of contextual informatDm canl~.t 
solve all of the prolflems of art.Mes, it is hoped that at 
least in some eases incorrect possil)ililies (:;m I)e elimi- 
nated (following tile "l>est guess" policy). I1, lhe cases 
that the context monitor cannot decide, an article, the 
MT system default will be re.lied upon. 
% decide between a definite and indefinite article in 
\]'~nglish, a simple rule of thumb in the present system 
is that once an object has been specified in a coiitexL, 
6ttSstlnlillg that the noun is defined as phu'al l)y sonic (~lher 
process, oLherwlse a vtdeo is als~ a l)osslbilily 
;ill subsequent references to that parLichh~r object in 
the saine context will be definite r. 
In the method proposed here, as objects are anal- 
ysed, they are giwm a unique reference number (re*) 
that separates them l'ron~ all other objects of file same 
type. Thus, the first time that an object is analysed, 
it will be nLade indefinite, unless the reference can l)e 
analysed as being ;t generic one (e.g. The lio~ is a 
da~lerous animal et(:). 
l"rom then on, if an item ill the text can be linked to 
an il.em which is the current focus, a potential focus 
or an i(.em on the focus stack, it will l)e made definite 
i. the English tr;mslation. Therefore, the two video 
models of Sentence 3 are recognised as a sul)set of 
the four videos that Form the focus and are. given the 
(lefild/.e article. 
Note also that as subparts of objects are inch|ded in 
the features attached to selllantic items using the has 
feature, objects related to an item already ment.ioned 
call also be treated to solne extent and translated with 
deft. ite a,'tich.'s: 
llannkr) bought a new video. She I.ook it 
back l.o the she 1) as II~c gape. head was 
damaged. 
This, howew% a very simple apl)|'oach nnd cannot 
accounl. \['or :Ill possibh', uses of the definite/ indefinil.e 
articles, lh)wevm', the appro;u:h outlined above also 
t~llows the "besl ~j,rss" strategy; where, this strate.gy 
fails the nornml (h:l'ault rules of the translation system 
Lake over. 
4.6 Restrictions on the Repetition of" 
\])roltouns 
In l';nglish, ov,'rt pronouns are repeatabh~ and in sonic 
cases obligatory in a sentence to preserve meaning. In 
,\]~ll'*g/llege ,however) overt pl'()tlOllllS arc llot repe~/tal)le 
as shown ill the i)elow s. 
lie dtx'.s his work when Jig wanls to. 
(he ,...;~ he..a.iwanls to when his worko~jdoes ) 
himself ~ 
~)~':{J: (b I.k:v, L- ~N ~,:-_ I;I,@a)41:T\]!:@- -~.7., 
his nwn 
It is; t.herc:foro. ,Jesh':fl)h'~ t.o have a routine in an MT 
system to rel~lace ow.'rt pronouns in English with 0 
or I'l~a v ('jib,.' oneself) in Japanese. In this case., 
the use o1' the p,'onomt lie in English will he analysed 
and recognised as rel>rring to the same person using 
r'rhls basic i)rlnclpll! is, supple~nent.ed I)y rules based on syn- 
{;ic\[ic i;OllS( i'llcLithllS t!{c 
a'l'his exal,q)le I;tken frt,.n \[\Vada 90\] 
the processes outlined above. Separate rnles concern- 
ing co-oecnrrenee of pronouns can then be used to 
substitutc ¢ or \[~I~,\]" ('jibun' himself) in t.he Japanese 
translation. 
5 Limitations g~ Problems 
As shown above, the inferencing carried out is very 
simple. It depends entirely on the links between nodes 
of the network and there is an obvious limit as to 
how complicated those links may I)eeome before the 
processing required to search all t,he nodes linked to 
a particular item becomes prohibitiw~'. At, the current, 
stage of planning, a strnctnre (a semantic il.eni) may 
be linked to another via one node (constrained l o be 
an Abstract It.elation). There are no current plans I.o 
increase the number of such linking nodes. 
The inference mechanism is also expected to per- 
form poorly where actions denoted by a wn'b are COlll- 
plex. This is due to the very simple feature descrip- 
tions that we use in the systenl. It might therefore 
be desirable that, if the processing is not ccmlpleted 
within, for example, a constrained time, the process 
be terminated and the context moiiit,or left to rely on 
semantic feature matching alone. 
Another major problem is writing the features for 
the links in the network. At the moment, all features 
are written by hand, })tit, it is hoped that sinlilar in- 
formation might be extracted from semantic and case- 
fi'ame dictionaries. 
The context monitor is currently written in PII0- 
LOG 9. The program eui'rently consists of several hun- 
dred lines of PROLOG. 
6 Final Remarks 
The idea of using contextual information in Ma- 
chine Translation has been proposed before (\[br ex- 
ainple \[Wada 90\],\[Elwrle 92\],\[llaendt 921), ho,vev,'r, 
there SeelfiS to be little researc\]l carried ()lit ill the 
field. MT res0areh still seems t,o take I.he sentence as 
i.he basle unit of translation and I.he qualil.y of thc.ir 
raw outl)ut snft'el's as .% restllL ~,Vc have prolmsed how 
some of the errors of J-li; & 1;3-21 translation can I>e 
solved mid have outlined it Context Monitor with shn- 
pie inferencing. 
The best guess apl)roach tries to define a l)rol)leiil 
and specify the informatioa needed to solve that prob- 
lem. The context monitor syst,em searches tbr Sl)e('ific 
information fi'om the input sentence and if it: cannot 
find it, it simply does nothing, ailowilig (.lie defaults of 
the translation system to supply the necessary hlfor- 
mation. The search routines of the context monitor 
look for that specific information at as earlier a sl.age 
as possible in the process aud so if that inforniation is 
9Not all of the feahn'es inentioneJ, in this paper at'(! currently 
iniplelriented 
not fonnd, the next routine is tried as qui\[:kly as pos- 
sible in order not to decrease the overall translation 
speed by a significant amounL 
Even when the context monitor fails and the MT 
systeln defaults are relied upon, the context monitor 
enstlres COllSistellcy with snl)sequent sentellces. 
Complicated texts are likely to lead to the Context. 
Monitor' failing often but i(, is still I)IL that the bet- 
ter translation produced iH many more cases and the 
fact that interference with the speed of the translation 
is negligible mean I.ha~; the prospects for a coral)act- 
sized personal MT system producing better quality 
translations are very promising. 
References 
\[Amores 89\] 
\[Eb,wle 92\] 
\[llaem:I( 92\] 
\[lsahara 86\] 
\[Isahara 87\] 
\[Ishizaki 89\] 
\[Isahara 90\] 
\[Quillia,~ (;3\] 
\[Sidner 81\] 
G. Amm'es, DIS(:. l)isseriati(m, UMISq', 19S.q. 
I(. Ehcrle, W. I<aq)cr, C. \]/ohrer, G'ontcx- 
lltal Conslrctinls for MT, 'I'M\[-92 l>roceedings, 
pp213-22.1. 
1(. llaenel(, Towards a Q)talily lm'proveme'nl 
in Mrchine 7'ra'n.Hation: Modelling Discourse 
Str~clure and lnchtdin9 l)inco'ttrse Develop- 
"merit i7) the I)etermim~tion o\] 7'ranslation 
EqMvalcnts, TM\[-92 Proceedings. 
I I. IsMl,:tra & S. Ishlza, kl, (/onte~:t Analysis Sys- 
teTn for Japanese Te:r/, C'O1,1N(\]86, 1>1)24.1-2,16. 
II. Isahara, K. llashida and S, \]shizaki, Co*t- 
slrt:iTlt Based l~fcre~ce /or Concept I/nification 
i~t Contt:et~tal A'a,tl~sis (in Japanese), I'roc of 
351h Meeting ()f the inf,,rlnallon Processing So- 
ciety of .Iapan, 1987, pl:,l 2,10-12,19. 
S. lshizaki & II. lsahara, Contea:tu.I In/of 
marion Tra'ttshtlio~t ,%sle'm. CONT'I~A.S'7' (in 
Japanese), 
JOtll'll&l Of the hifi)rmati.n l)rocesslil,g S()(:icly 
~al~.(}) no. HI, l)l)l.lOS-6) I(089. 
H. isahava ,~z .'-;. Ishizald, Natural Lan.t:v.a#e Un- 
derstandbl9 b'vslen~ wit\]t (\]onceFl Ilierarchy, 
Prec. of l>acific lllm Internaiimlal Conference 
on AI '90, pl)322-27. 
M. IL Quilllan, Se'mantic Aft"m- 
o)'V, ill .N(!tllaudic hffnrmali.nl Proces~.ing, cd M. 
l~.linslo', i'.)(;S, pi)227-270, 
C. I,. Sid- 
tier, \]"oeusl))g for lnlerp'relallo'n of \])eo'aolt'ns, 
Anlerican Journal ()1" (~OnlF, utationa.l l,htgulsllcs, 
\[$idner 80\] 
\[Wa,la 90\] 
Vtd.7, 217-213. 
C. I,..qldn,n', I"ocusin9 i~*. the Comprehensio~t 
oJ Definite Anapho'ra in 
Headings in Natural \[,a.'nguage Processing, eds 
l~..l. (h'.~,z, I(.S. Jones, B.I,. \Vebber, L986. 
II. Wada, l)isco~rse Processi~t(i in MT: Prol,- 
lems in l'ronomi'nal Tr~tmshttion, COI,1NG-90, 
pp7:L75. 
56 
