Con, tructlon of a \]3ili:ngua, l )" • 1 tci, onary 
Intermediated by a Third Langua, ge 
Kumiko TANAKA Kyo.ii UMEM\[JID\ 
\])ivisim, of Engineering, University of Tokyo. NTT Basic \]h, sea.rch \],almratm'i(,s. 
7-31 Ihmgo, Bunkyouku, Tokyo, 113, \]la.palL 3-9-II Midm'i, Musashino, 'l))kyo, t80, JalmU. 
kun~iko(gipl, i;. u-Lokyo, ac. j p umemur a@nue ram. NTT. JP 
Abstract; 
When using i~ third l;mgu;~g(' to const:rl,(:t a bilin- 
gum dictionary, it is necessary t;o discriminate 
equiwflencies from imwprol)riat:o words deriw~d 
as ,~ resnll; of ambiguity in I:hc third lat<gmtge. 
We propose it metho(l 1:o t;re,d: this by utilizing 
the strucl;llr(m of di(:donari(,s to measm'(~ die near. 
n(,ss of the meanings of words. Th(' resuli:ing dic- 
I;ionary is it wor(l-I:o-wor(\[ })ilingu;d dictionary (,f 
nOllIiS and can })e used t:() refin,~ I;h(, (mt.ri(m and 
equlvalm~eies in published bilingual dicl:hmarie'.. 
1 Introduction 
When vocabulary cmmot be found in bilingual 
dictionaries, it is freqnendy obt;ain(;d by using a 
l:hird language as an intermediary. '\]'his imlical;es 
thai: SUl)i)lemei~tal informal, ion llla, y lie in ot;her 
forlns in other dicl;ionarie.~< I\[(;re we Lry nsinp; 
electroni(" (ti(:i;ionaries which cau t)e rel'ormod (m 
it large s(:ah.', 1:o extract; this iMornud;ions so thai: 
w(' can ohi;Mn subsidiary (la, t;a and re{hi(! a (llr(~(:t: 
bilingual dictionary. 
Looking up words in bilingual dicl:i(maries in 
termediating the third llutguage is a inethod ofl;en 
used by t)eople who handh.' m~(:omm(m \]itll,~ll;14~(;s 
in a specific domain. If this i)ro(:ess c;m l)e mlt;O-- 
nutt(,(t, biliI~gual (tiedonmries of any kind l)el;ween 
any languages may l)(: ol)taine(\[ its long ;m l:ho.'m 
concerned languagos haw' di(:donaries to a (:(mt- 
mort hmp;uage. One objecl;iw,. (ff the r(!sem'(:h r('- 
l)orte(\[ here is to est;al)lish ;t first st(' 1, in itlli;OllH'ti,- 
lug t;his i)roeess. 
To consl;ruet, a £1apanese~-+lh'ench dictiom~ry, 
we chose English its tim inte.rme(liary l;m/,;u;q,/(', 
beta, rise JN)anesev-+English and Englishv-d)'rench 
dictionaries e.xist in electronic forms and because 
publMw, d aN)anese~l,'rench dicl;ionm'ies provide 
enough vocabulary in comparison with the result- 
ing dict;iona.ry. 
In Section 2 we describe a method for exl:ra(:t;- 
ing equiv,'denei(;s for a given wor(I. \]l;s fmMamen- 
tel concepts m:e stated in Secti(m :3. Tim whole. 
procedure used to construct th(', new dictionary is 
shown in Section 4 ~tn(l in Se(:t,i(,n l\] t:he r('suli;ing 
dictionary is evaluated. 
aapml(,se-English, English-J~tpltilese, Euglish- 
Fren(:h, French-F, iw;lish , a;q~mmse-l,'renc\]l and 
Freneh-Jal);mese dictionaries are rest)eel:ively de- 
not:ed l)icj _~, Oico._Q, Dice_ ,:e, I)ic:e_ ,o, Dicj. <f, 
and D:i.c~. ,j, Dic .... 7J i'~ called ~m i'n.verse die-, 
t'm.na.r!\] of Dicy ........ J:q)atm.qe words \]uw(; hfl'or- 
mat:ion in the followi.t,~ \['ormat: promtn('i~tl:ion in 
r(muji, and i1..~; ('(lIfiwd('at(:c in English. English 
words are writ:ten in this lbnt and Frenc\]l words 
in this fonl;. 
2 Ov(u'view of the M(~I;ho(l 
2.1. hlverse ConsulI;ation 
'.l'he m(,sl> naive w;~y t:o use l,\]ngli:di to ol)tain 
l"reH(:h wor(ls (:(~rres\[mB(liltg I;o :t ,\]a,l}:Ul(:s(~ w(n'(l 
is t:o h)ok ltl) t\]l(~ Jalmmese word iu a Dicj_.e 
am(I th(m look up {:he rosultaut; En,,~lish words in 
;~ Dice ,/,. The r(m,dlting Fren(:h words (:m~ I,(: 
regard('(| its e(luivMenco c~mdldates (E(\]s) o\[' the 
origin~d Jal)an('.:m wt)M. For ex;ullple, ill Fit';. I, 
\]",:Cs for a ,lal)aues(, word *'~:*'" " ~,~\]~ (kyousou : (:,m- 
pel:itiolO are comp(.'i;ition, contours, race cix:. 
Am()nl< these, race a, nd hSte are ina(h.'(lUatta~ a,q 
,', ,Z \] "1 O(ltfivahufl;s of >~.l . 
As for rat(;, the. l'h@ish word race has several 
m(,aninv~s with l:he sam,, slmlling: otto is to co're.- 
pole and another is h, wma',. "rata. Ig is h.',,ma'n r(u:c 
which induces the inad(~quat;e EC race, As for 
h/it('., the English wor(1 race has the widm" m<uv 
ing to h'u'rry whMt the (n'it~inal Ja,plmese wm'(l "}?~{ 
(fJ-" does not. Since hSte is i~ dire(:t t;ransl~tt:iou 
of to /vurry, i|; is in;q)propri:~:e as an equivahm(:(< 
The fifth)wing l:lm,,e (:ase,~; l,;eneral:e irrelewmt: 
ECs. 
I. An lih@ish wor(I wit.lt t:he same si)ellin/,; 
lml; wiLh <lill'erelfl: nlemfings is inl;erme(liated. 
(race in t;he al)(we ('xamph.') 
2. An English word wil.h a wider IneiLl,ill/Y li\]lttll 
that of originid .lal)mW.:;(! wor(\] is intermedi.. 
ated. (h&te in the Mmw! exam.ph!) 
3. '\])here are misI;akes in dict.iomn'ies. 
The lirsl; two (::tses ~tr(.' dlle |,o t;he ambiguity in En- 
tJish. An English word with ;t narrower mo.~mln/,; 
I;hall the ,\]itl)a, lle,q(', SOl/l'ce lna~y miss ,qOlll(~ French 
equivah;nl;,q. W( 3 think thai; il' du.' origimd word 
has ambiguit:y a, nd several meaning,, I;he (licdo- 
nary gives t:he (:orr(,si)(mdinl, ~ English wor(ls. 
We couhl han(lh~ the l~rol)lem of choosing equiv-. 
alenci(~s t'rom ;tltlOllg \]'\](\].q by :{elll~tll\[.i(; pro(:essin/,;, 
but expr('ssing m(;a.ing of vocMmlary in (lieti(,- 
nm'ies is a /,;r(.'~tt l)robh.'m. A simpler way is I:o 
look t;he ECs u I) in l;he inverse (li(:t,i(mary. For ex- 
toni)l('. ~ (,he (:ml consult Dic, ,j for COml)6.tition , 
contours, a u(I rac(', a, ud du;reby gel; ~9, J , De, 
297 
Fig. 1 
Japanese English French 
compelRIon -- compefitilm <- 
_ enn,.s, .... 
raco ~ CO rse 
Equivalence eandhlates (ECs) for "~'. 
~ comp~titim~ ~ A 
(Selection Area) 
~ rnce 
Fig. 2 One time inverse eonsultatlon (\[C~). 
flY", and "&}i~" (jinshu : human race) as their 
respective equivalencies. Since ".&~i\[(" has notl,- 
ing to do with "~(/,'", race is excluded. We 
call this inethod of looking up ECs in the inverse 
dictionaries when choosing relevant equivalencies 
inverse consultation, and we call the words ob- 
tained by looking up inverse dictionaries the se- 
lection area (SA). Inverse consultation utilizes the. 
structure of dictionaries to measure the nearness 
of the meanings of words in (tiff>rent languages. 
The simplest application of inverse consultation 
is to use Dic~_,~. In the above example, each EC 
is looked up in Dict_,e and the results are com- 
pared with the English equivalencies of ")~'6"" (E); 
namely, competition, contest, and race. The SA of 
eompdtitlon is competition, contest and match, 
which have tile elements contest and competition 
in common with E (Fig. 2). As compdtition 
derived from competition, competition should be 
put aside, but contest is still left as a common 
element and thus compdtition is selected as an 
equivalence of '"~(P"'. As for race, the SA of 
race consists of race and ancestry, whose inter- 
section with E only gives race; so race is judged 
as an inadequate EC. In shorl;, the number of ele- 
ments in common between the selection area and 
E indicates the nearness of the meaning between 
the EC and the original word. 
For the inverse consultation (lmseril)ed above, 
the SA was in English. If we use Dice~j as 
an inverse dictionary successively afl;er consult> 
ing Dic~-.e, then the SA is in Japanese an(1 we 
~f\[' with the SA (Fig. 3). The. SA for eonlpare "'~'~ "" 
SA 
~ contest -~_ 
~ ¢Omlletition compet~ 
malch 
race ~ race 
ancestry I~'~I 
Fig. 3 Two times inverse consultation (IC~) . 
compdtition consists of two JeJt ~( s an(1 three 
"~j~5~"s (kyougi : game). For race, the SA has 
J ~,\[,. , "~ld~ll.' (s(nzo : ancestry). Siuee ,,~ff,.,,, ,, ~. ..... 
~e.=T(~ aI)I)ears only once for race, we (lis(:ar(l 
the. EC race. 
There can be a infinit.e number of inverse (:on- 
sultations according to the nmnber of (:onsulte(1 
inverse (tietionaries. If the inverse dictionaries arc 
consulted 'n times, we call the method n ti'mes 
inverse cons'ultation. (/C',,). Which inverse di(:tio- 
nary to use. does not always have a unique answer. 
For IC2 for example, we ntay consulU Di%_** af 
ter consulting Dicf.-.e with the SA in French. 
2.2 Selection Procedure 
Once the SA tbr a give.n woM is obtained, equiv- 
Mencies are selected by handling two collections 
of words. We (:all this process the selection p'ro- 
ced.ure. 
One way to do this is to count the mnnber of 
specific element:s in the SA. For example, if the 
SA is in Japanesc', the number of l.he element ""~¢~d 
@"' itself is counted. Another way is to count how 
many parts of words (PWs) are eontMnmd in the 
SA. " ~a and "¢6" \[ or exaulple, the number o\[' ""'~:" 
contained in the SA is considered (thus "" ..... ~ in 
"~j~'~" iS ;dSO collnl;ed). 
If we handle the meaning of words, a third way 
.v)¢.~ ~ in ;t Japalmse Lhesaurns an(1 is to look up ' *'~-" 
COllar how nlal\]y times the synottynls ai)i)ear in 
the $A. For exanq)le, il' ,. is a synonym or .,%,k ~" 
, ~)t.)', , then the numl)e.r of at)i)ear:mces of "~j~ 
~( ~'lfi: z" I ~}:" is added to that of ~,g ~ . If we go further 
to h;mdle the meaning, we might as well process 
words by semantic processing. 
Since a kanji is an ideogram, the second method 
also handles the meaning. When 1;11o s(,qection 
area is in English or l"rench, the eorresi)on(ting 
me.tlm(1 is to count morphemes such as "inl;er" 
and "national" for the word "int0.rnationar'. W(; 
are interested in to what extent the method that 
does not (;xplicitly concern the meaning may be 
used to handle the me,ruing of words, lln |;11(; for 
lowing, we. focus on 1;11(; former two inethods. 
29B 
3 Fundamental Concepts 
3.1. Harmonized Dictionary 
A bilingual dictionary forms a grN)h whose nodes 
are words and whose })ranches are correspon- 
dences between the words. Branchc's haw; direc- 
tions, which make the graph asymmetric. D:ic.,(1,g 
is a graph with all the tn'anches in Dic> 'v in in- 
verse direction. 
Since the purpose of bilingual dictionaries is 
to denote the correspondences of words that 
have the same meaning \[tIar83\], it; is Iud.ura\] that 
branches ~re bidirectional. We t;herefore design 
symmetrical dictionaries and we denote, a di(:tio- 
nary from hmguage :c to y as D .... u ,calling it; ;t 
harmonized dictiona'uI. When D,._ ,,/and D:j ,~, are 
construc~;e(1 fi'om the same dictionaries, D!t.**=: 
DTly holds. We remove the overlaps of branches. 
3.2 Syntactic Selection Procedure 
A multiset here is a set in which each element: has 
:~ weight that is a m~tural nmnl)er. The weighl; of 
ml element is defined as the nmnl)er of times it 
appears when looking up words in dietiomwies. 
In the example shown in Fig. 3, the nmltiset SA 
~\[(' with weighl; 2 for eompdtition ('onsists of "~"~ .... 
and ")j~J,hi" with weight 3. We denote the. weight: 
of eh;nlent x in multiset X as ba(X,x); for in- 
stance, 5~(SA, ~e)~-Tf~**~ ) = 2. Using the same nota- 
tion, 5a(X, Y) is defined as ibllows when X and 
Y are multisets: 
~'a(X, Y) = ~ 5a(X, .,) 
yCl / 
When a multiset Z consists of" ,,,.,,:zw)~.)q ...... and ma~z,~'~"" 
the.n G(SA, Z) = 5. 
The nol;;tl;ion ~b(X, X) represents the sum of the 
weights of the elements that contain PWs of :c 
in Inultiset X. For instance., if PWs are defined 
as k~mji, 5 / ..... ~a, bLot*,~.~q) is 7 by adding 5 (sum of 
weights of elements in the SA that have kmt}i 
"~") ~ul(l 2 (SlllII Of weights of elemenl:s ht I.he 
SA that have kanji "q""). Using ~lw. same nota- 
l;ion, 5a(X,Y) is defined as fi)llows when X ~md 
Y are multisets: 
eb(X,V) :: ~ eb(X,v) 
y~Y 
For instance, 5b(SA, Z) = 12; that is 7 plus 5 (=~b(SA,'~)).We 
,,se the .,,t~|:i,,,, ~ .s ~ >- 
ramet;er for ~a or ~5 b. 
3.3 Properties of Inverse Consultation 
In the following, we use Df.~ e ~tn(\[ De__+ j ~ts hlverse 
dictionaries when starting fl'om Japanese words 
:tnd we use Dj_~ e and De_~ f as invers(,, di(:timmries 
when starting Dora French words. French gCs for 
a Japanese word j form a muir|set F expressed as 
g = D~-.2Dj-.~j • 
I~'c.nch equivalencies selected })y fC1 forltt a 
Japanese Ellgllsh French ~ 
o ...... o 
o o ~ Irrelevant EC 
Fig. 4 A structure that £, is inapplh:alde. 
multiset: whose e, lemenl: f satMies 
f ~ F mid 5(Df-,ef,Dj--~ej) 2> 1 . 
French equiwdencies selected by IC,~ form a mul- 
tisel: whose element £ s:tl:isfies 
f (~ F ;tIld ¢5(De ~jDf-,ef, j) > 1 . 
In l~he following, we focus on the ICI and 1C2 
described above and ex:tmine their proimrl;ics. 
Property l 
5a(De- ,jD:f- ,c~, j) -- ~a(Dj-~ej, Df~ef) 
= 5a(De_÷fDj_,ej , £) 
This properl;y indicates that when using selection 
procedure b,, equiwdencies selected by ICI ~tn(\[ 
I(2~ ;we. exaci,ly the stone. Moreover, it; is suf- 
ficient to choose ECs wltose weights are gre~tter 
than I in F. '\['he proof of Property \] is shown in 
Appendix. 
Property 2 If5 is 5,, the res'nlting ,\]apa'nesc- 
French dictio'na','y is a harmo'nizcd dictionru'y. 
This is .not al,~ays true fo'," &,. 
This properi;y is cle~u' with symmel,rically strut-. 
turcd dictionaries. When St, is used, t;he resuli;inl, ~ 
dicl:ionary depends on how PW is de.fined and 
doe.s not ~dways l)ecome a harmonized dictimu~ry. 
Property 3 If there is (t st'ructure s'u.ch as 
shown in Fig. ~, 5b must bc used as 5 to e:rcl'ude 
ir'rclcva'nl ECs. 
When 5 a is used, t;he SA for two ECs will be ex- 
ael;ly the same, which makes i~. intpossible to dis- 
card inapprol)riate ECs. All.hough this kind of 
structure seems t;o b(.' rare, it cmt exist lw.(:ause of 
the hist:m'ical transition of words. When a sit@e 
l);nglish word is int.ern|edi;d:ed, inN)ln:oi)rial;e ECs 
cannot be discarded by using 5a \[or ICI. 
4 Experiment 
4.1 Dictionary Data 
The dictionaries used in the extmrin~ent ;u:e 
Dicj .... \[Ich9II\], Dice~ j \[I(oi90\], Dice_,y \[For82\], 
m~d Dicf_ ,e \[Led82\]. The whole exl)erimental pro- 
cedure is shown in Fig. 5. Word-to-wm'd dietio- 
nm'ies are first extracted fi'om each dictionary. All 
words ~we nouns; in I~m:tic,dar, they are one word 
n(mns in English and \]"tenth. Since the diet;|o-. 
mu'y synt~tx was not; :tlways consisl;ent, word-to- 
word dictionaries toni;a in some mistakes (imuh> 
quate CO1TeSpOII(\[(!IICeS), 
Ilarmouized dicl:ion:zries are thou consl;ruete(l 
from the. word-to-word dM:ionaries as follows: 
299 
French-English Dictionary 1 L-- Dct°naryEnolish'French I 
Word-to-Word I I Word-to-Word I 
t apanese-Engl'ish'~ English-Japanese Dictionary | Dictionary 
Japanese-English English-Japanese 
| Word-to-Word I I Word-to-Word | Dtct~on~nary 
apanese-English French-English nglish-dapanose English-French 
Harmonized Harmonized D,cttona~¢ \[ D ctionary 
I Japanese-French French-Japanese 
Harmonized , Diclioneq/ 
Fig. 5 Whole procedure. 
Dj ~o = De-lj 
1 De-.f = Dic~eU Dico-~f 
Df-~e = Do._.+ f 
Although there are other ways to symmetrize dic- 
tionaries, (for example, by removing all branches 
that are not bidirectional), we chose the above 
procedm:es for the lexieographical reason de- 
scribed below \[IlarS3\]. 
There are two kinds of bilingual dictionar- 
ies, one is from ,~ foreign language (f) to the 
nmther language (Die/ ..... ), and tile other is fi'om 
a mother language \[m) to a foreign language 
(Diem_,/). In Dic/ ...... when there are no equiv- 
aleneies in m for a foreign word, the dictionary 
gives its definition or explam~tion of the word in 
m. Therefore, all the foreign words can be con- 
tained in the dictionary. In Dic,r>~f on tl,e other 
hand, if there are no equivalencies in f for a word 
of m, the word itself is often dropped fl'om the 
dictionary. The words contained in the dictionary 
are therefbre a part of m, and Dic ....... f lacks many 
entries. A harmonized dictionary is a solution t;o 
this problem because it contains equiwdenc, ies of 
Dicf_.,~ as entries of Diem--,\]'. 
4.2 Procedure of Inverse Consultation 
lh'om Property 1, we use 5a wi~h IC1 and we use 
5b with IC'2. The PW for 5b are defined as for 
lows: 
• Japanese 6353 kanjis. 
• French morphemes \[Mauaa\]. 
1151 prefix and 710 suffix. 
ih'om Property 2, inverse consultation is applied 
to both Japanese and French entries, and then 
tile results are put together to construct a harmo- 
nized dictionary. We denote it as Dj~t or Df_+j. 
Each entry within Dj-+e and Dt--,e is classified 
into one of five types according to the procedure 
used to select its equivaleneies from ECs. 
• Type A A single EC exists and is selected 
unconditionally as the eqnivalence. 
• Type 13 Equiwdeneies 1)y \[CI exist; all are 
selected and the rest of the ECs are discarded. 
• Type C There are no equivalencies by ICt 
but there are by \[C.2 . One third (empirically 
decided) of ECs by 1C2 arc' selected according 
to the wdue of the following fl'action (larger 
ones are seleet:ed): 
Value of flmction 5b 
SA (byte) 
• Type D No equivahmeies by ICt nor by 
IC'2 appeared but there are several ECs. For 
this entry, it is impossible to select the rele- 
wmt equivalencies. 
• Type E There are no ECs. 
Entries of Type A acquire more appropri~te 
equivalencies than do entries of Type B, and en- 
tries of Type B acquire more appropri~d;e equiw> 
lencies t;han do entries of Type C. 
5 Ewduation of Experiment 
5.1 Result of the Example 
Fquivaleneies ol)tained fin" the Japanese word "'~ 
• .~1." are 
eoneours, rivalitd, comphtition, 
eollrse~ concurrence, 6mulation 
and the intermediated English words are 
competition, contest, rival, rivalry, race. 
The entry ";~(/'~-" is of Type B. '.the, munlmr of 
ECs is 4\] (inchading overlaps), and 13 of thenl 
are selected by ICt. The number of ECs in each 
category of irrelewmt words described in Section 
2.1 is listed in Table 1, which indic.ates that in- 
verse consultation can detect the relewmt words 
even when there are ndstakes ill the dictionar- 
ies. The word which should not be dropped was 
joute. 
Equivalencic's for ,.,,~.z. in \[Tam85\] D~ P are 
300 
Table 1 Details of ECs of "~SF". 
Different meaning in English 9 
Memdng extended in English 
Mistakes of dictionaries 
Words which, nmst, not be droplwd 
Table 2 Classification of entt'ies. 
D j_ ,a" 
Total 42190 (10(}.0%) 
Type A 1600 (3.8%)) 
Type B 8179 (19.,1%) 
I 'l'vtw. (d 24(),17 (57.0%) 
I 'l.'wm l) 1514 (3.6%) 
I 't'vt)e E 6850 (16.2%) 
0, ,j 
~,37m (1oo.1~%) 
38~;~ (1~,~%) 
7397 (31,3%) 
6,t52 (27.2%) 
~988 (\]~.(;%) 
30:!\] (~2.7%) 
Table 
eategory 
Cultural words 
Technical terms 
or l)roper ll()llllS 
Borrowed words 
Entries witt, no FC. 
examph* l';n/~/lish 
cqttivalenee 
:t3 :~F .-'\],{ (otoshidama : handsel 
Our tradition to give 
mone.y to childrei~ on 
New Year) 
c&lille cedilln 
~f')' .X (gwns : gauss) gauss 
Cicdron Cicero 
~y..e. U q7 (al)eriehihu aperitif 
: apctizer) 
tee-shirt teeshirt 
coneours, rivalit6, competition, 
eollrse~ cone.llrrenee. 
Our result contains dmulation (which means r'i- 
yah'y) in addition to the entries lit the published 
dictionm'y. 
5.2 Evaluation of Entries 
Table 2 lists the each number of enl;ric's 1)elong-- 
ing to the Tyt)cs A~-,D ((lefined in Section ,1.2). 
'.12ype 1) consists of the following entries. 
* One English word is inte.rm(,.diated ;rod sev- 
end ECs apt)e~u'. 
,, The entry contains no PW. 
The Type D percentage difference betwc(',n Dj__,f 
and D~_,j shows that the nmnl)er of m1`trics of 
Type D det)ends on the numl)er of registered 
PWs. 
No EC appears when ~m English word t:o be in-- 
termediated does nol; exist as the curry of De ,f 
or De-,j. Such entries can be categorized as in 
Table 3. 
Original words are apt to t)e transl:ttcd into m> 
common English words, so they normally do not. 
appear as entries if the same kind of words (la not 
exist in tit(,' objeetiw', hmgm|ge. '\]'echnic~d t(,rms 
and proper nouns det)end very much on eultm'e. 
q'alfle 4 Ewduatlon of equiwdencles. 
__ D-JZ'~ _A' "-'~I 
rate I'I1 \] l(2 R1 I 112 I 
80%,-,100% 58 I 56 18 I 58 I 
60%~80% ,i, i 1..1, 8 I 15 I 
,1{}%,-{30% 9 I 13 13 I 9 I 
20%~,10% ,u I 10 24 I ,1 I 
()%~2()~, '22 I 7 37 I M \[ 
Many French 1)lacen~m~cs, for instance, arc 'l'ypc 
F,. lhn'r()wed words ave expressed in inconsistc,lt 
spd\]ings, mid which o\[ l;hem are to be found in the 
dictionary ~dso depends ou I:he cuitm'e (apdritif i.'; 
t:he eqldvalence in D±cj_,~.). 
Since ~ hm:monizint,; dictionm'y mtgments the 
entries, t;he resulting dictiomtry contains entries 
thaL are not in the lmblished dictiomwy. As ex- 
plained in Section ,l.\], this phenomenon is con- 
spicuous when we compa, re Dj_,f with the pub- 
lishcd dictionary \[shtz70\]. '.l'he,,', entries can be cati:-- 
egorizcd as follows: 
1. Colloquia.1 words. 
Ex. ~)/vl~ ,p/,~ (wanchan : ImPl)y) 
2. Tecllnic~d terlllS or \[)rol)er ,to,,Hs. 
la;x. y X-'VS,', 1" (~tSltbeStltO : aSl:,CStc') 
3, (JOlllpOll,ld IIOllltS. 
Ex. )K{\[;{g')I\] (lmukasayou : disinl:cgrate) 
!1 !/J 1\]11 i '~: 51" i\[~ .~,~ (buld~uml:eiseisaku : 
wdot'izal;ion) 
ilarmonizing dictionaries help t;o g~,thcr corre. 
spondctlces between the inotltcr language ~md at 
fln'eign langu:~gc' and ~n'e usefitl in revising pub- 
lisht,d dictionaries. 
5.3 Ewduation of Equiwdenc.ics 
We ewdmtted the equiwdencies of resulting dictio- 
nm'ies I~y con|p:u'iI|/,; I,ltenl with I;h()se o\[7 lmblished 
dictionaries \[Ta,,85\] \[S,zT0\]. For vm|do|n 1()(1 en- 
tries in both dictiona.ries, the following two imr - 
c(!ul;n.gcs ((:/d(:ul;tl,cll mn.nlmlly) ~u'c list{xl in 'I';d>h'. 
el: 
® RI F,'aetion of eqldwde|mie.s in the lmb- 
lished dlcl;ionary which were also found by 
this method. 
• 1{.2 l"raetion of equlvalencies found by this 
method which were judge(\[ ~q)prol)riate. 
Note thud; entries with greater R2 cont~dn N)i):'o.- 
pt'iate c'quiwdencies on higher rate. This is not 
true for R\], since 11.\] indical:es the discrep~mcy 
of equivalencies bei:wec'n the resulting dicgionary 
and published dictiomu'ies. 
Enl,rics of 112=\]00% ca,n 1)(,' classified as follows: 
1. Entries of '\['ype A or B. (See Section 4.2.) 
2, Entries l;ll~d; \]taw,' less \[.hart three \],;Cs. 
Fmtries with less than 3 ECs a,l~loltnt to 29.0% 
for Dj ~, and 34.2% \[k)r De- ,j. This indicates I:hat 
words with specific meanings are apt to ;u:quire 
~q)I)l:opriate equiwtlcncies. 
301 
Table 5 French entries whose R1 are 0 %. 
entry 
dlstique 
pull 
boulier 
Df~j 
ourenku, tuiku : 
the terminology for 
Japanese and Chi- 
nese poelns of sanle 
kind) 
~- Y-" ( seetm'~ 
: sweater wrltten 
with Japanese let- 
ters. This word is 
conllnon ) 
gfl,~ (sa.nban : Aba- 
cus in general) 
Published dictionary 
\[Tam85\] 
2 ~1: ~} (2 gyou-shi : coinage for the term 
for European i)oems) 
7/ )!/ :t-- J';~ (lmru 
obar : frail-over writ- 
ten with aal)anese 
letters. This term 
is not eonllnon ill 
Japan) 
(soroban, kazoedama 
: Japanese abacus) = 
Table 6 Japanese entries whose I1.1 are 0 %. 
entry Dj ~:f Pul)lished die- 
tionary \[Suz70' 
~0 <" g* (tsugumi : 
thrush) 
'~" N (share : wit- ticism, joke, jest, 
pun) 
grive 
astuce 
badinage 
drblerie 
facdrie 
farce 
plaisanterie 
merle 
calelnt) our 
The entries tend to have either R1=80%,,~10{1% 
or R1=0%~,20%. Of the latter, some examples 
are listed in Table 5 and 6. For distique in Ta- 
ble 5, which is a term in French literature, De~j 
translates it into the term for the corresponding 
kind of Japanese literature. Although it resem- 
bles to the direct translation of distique, it is 
only an analogy. On the other hand, since the di- 
rect translation of distlque in the published dic- 
tionary adopts the concept of French imems, it 
is uncommon and cannot be mttlerstoo(1 \])y iilOSl; 
Japanese readers. The same is true for boulier 
except that the commml Japanese is indicated in 
a published dictionary. Pull is borrowed front the 
English word pull-over whose direct translation is 
contained in a published dictionary, and it is not 
a common word in Japanese. 
In the first example in Table 6, grive is the 
generic name equivalent to th rush, whereas merle 
is a kind of thrush. The second example shows 
that more equivalencies are found in Dj_d. 
To sum up, the resulting dictionary can be uti- 
lized in conjunction with the published (tictionar- 
ies as follows: 
• To revise the equiwdeneies. 
• To supplement the equivaleneies. 
Sections of the resulting dictionaries is listed 
in Table 7. List 1 is Dj_~, and List 2 is D,+j. 
For each list, entrie.s arc in t;he first row and their 
equiwdeneies are in the third row. Symbols in the. 
second row indicate how approI/riate each equiv- 
alence is. (Refer to t, he notes beside.) 
6 Related Work 
The use of a third hmguage English as an inter- 
mediary in the construction of a bilingual diet:to-- 
nary was test;ed manually on a large scale on c(lit- 
ing the Spanish-Japanese (lictionary \[Kuwg0\]. It is 
now a representative middle-sized dictionary hav- 
ing a large quantity of informal;ion. 
Toktmaga and Tanaka \[Tok90\] tried t;o extract; a 
eonceI)tual dictionary fi'om Japanese-English and 
English-Japanese dictionaries. Although they 
used a concept similar to ours that is the graph 
structure of a dictionaryrelate(\[ to meaning of 
words, their t?ameworks and final product differ 
frolil ollrs. 
7 Conclusion and Future Works 
The i)rol)osed method for using a intermediate 
hmguage to construct a l)ilingu:d dictionary uti- 
lizes the structure of dictionaries and morphemes 
and can choose appropriate equiwdencies for most, 
entries. Comi)aring the resulting dictionary with 
publishe(l dictionaries showed that tiara obtained 
are usefltl for revising and sUi)l)lementing t.he vo- 
cabulary of existing dictionaires. 
To increase the accuracy with which eqniva- 
h',ncie.s can be selected, mistakes in word-to-word 
dictionaries must t)e corrected even if' our method 
may detect aI)i)ropri:~te equiwdeneies. One way 
to do this would be t.o use thesaurus to cheek 
whether the extracted eorresi)ondences are rele- 
wmt. 
Nouns were l;akeu into consideration in tiffs re- 
search, and the next ste I) will be to apply the 
proposed method to other parts of speech. We 
also need to estat/lish a way to handle compom~d 
words in European languages. 
A,zknowledge.ments 
We exl>ress o11l- grai;itude t<) Prof. tlideya Iwasaki 
of University <)f Tokyo r<)r his guidance on the<)ret- 
teal aspects and essential remarks on exi)eriments. 
We. are also thankful to Prof. Nigel Ward of Uni- 
versity of Tokyo tot" his lexicographieal advice and 
for his inwduable aid in shapillg our idea. 
\[Bog89\] Bogm'aev, 13. et al. (1989). Computa- 
tio'n, al Lc:cicog'raphy for Natural La'ng'mtgc Pro- 
cessing. Longman. 
? \[1i'i185\] Filh:nore, C. (1985). lrames and the se- 
mantics ot'mtderstanding. Quade~oai di Scrmvn,- 
tiea, Vol.6. No.Z 
\[Fil92\] Fillmore, C. and Atkins, 13. (1992). To- 
ward a Frame-Based Le×ieon. The se'm, anl, ies 
of .I~ISK and its Neighbors. Academic Press. 
302 
Table 7 
!-f-q" 
!7"--b 
! T-- K:,. 9X 
! T--l, II 'Y 9 
!@B~k 
Listl 
0* 
2 
2 
1 0* 
0 {) 
0 0 
0 0* 
0 0 
0 
0 
2 
0 
1 
0 0 
0 
1 0 
0 0* 
1 0* 
l) 
1 
0 
1 ()* 
2 
1 
2 
2 
2 
0 0 
CiI~ttllll(~ 
crini~re 
tifs 
gmuinerie 
malice chien-chien 
chienchien 
t, outou 
chienochlen chienchien 
t, outou 
\[%rc arclie 
route 
art; 
japonisant arbousi(,r 
(:OILer h~ 
ell(JenchelttOllt, 
fcrnieture inunobilisatlon 
platine 
s(,rrllrc 
trapilhm 
D.II liuld (! 
,drag&; 
|)rOlllOtollr 
f(m(htteur discussion 
invite 
Imintage sollicitation 
v&itication 
rill) ~l.ll-C ach e centimbtre 
xnagndtophonc 
existence 
rouleau 
volute 
A part of resulting Japane, se,-~French dictionary. 
List2 
Zambie 0* Zai're 0* 
Zinllml)we 0* 
!Zorolc~tre 0 
al)aissenumt 0 
o 
o 
o 
o 
o* 
o 
o 
o 
o 
o 
o 
o* 
o 
I 
I 
I 
almndonnement O* 
0 almndou 0* 
0 0 
0 0 
0 
0 0 
0* 
al)aque (} 
/I 
l) 
0 
0 
0 
-Iy y ff "f 
"0'4 -- m 
"/r: TX Y-. 
aJIBVW \[,':lq I L, 
t,1~% 
a£;LI 
II%V Sliir/vk~l~'a)i!U I; 
;EP~ 
is\[; I" 
r;~it{ 
f~\[ \]" 
I!ll:~ 7 a)7,Ji I" 
/tliJll~. 
I~> :¢':'N 
~I!?R 
;{.,.,'fir: 
Jtz :1'~ 
3"<y 7 
Ni, t~ 
0: Judged apl)rol)riah? 
(counted fl)r 112). 
h Significance slighdy dif- 
ferent.. 
2: Significanc.e comphq;ely 
different. 
*: \[POlllld COIIIIIIOII ill I;WO 
dictionaries 
(couuted for II1). 
Mark 'T' af.tached to en- 
(:)'it's lllO\[lllS I;lllg Lhc en- 
tries were not forum in 
the published dictionar- 
ics. 
\[For82\] Forbes, P. et M. (1982). Shorter E'nglish- 
French Dictiona<q. f\[m:rap Limited. 
\[Itar83\] Itartmann, R. (1983). Le:6cography: 
principlcs and practice. Academic Pre, ss. 
\[Id~90\] Ichikawa, S. (1990). New ,lapa'n, cse- 
English Dictionary. Kenkyuusha. 
\[Ide93\] hie, N. and V6ronis, J. (1993). Exl;r~tctin<,; 
knowledge basis fi'om machine ri!adal)h,' dii:ti.- 
naries. KB~KS, JIPDIi3C. 
\[Koi90\] I(oinc, Y. (1990). New EngIish,-dapane,s'c 
Dictionary. Kcnkyuusha, 
\[Kuw90\] Kuwana, K. (1990). Spa'n.isl~-Japa'nese 
Dictionary. Shougakukan. 
\[Led82\] Led&err, M. et al. (1982). Shorter 
l,)'eneh,-English Dictionary. IIarrat) Limited. 
\[M;m85\] Maubourguet, P. et al. (1!185). G'ra',,,d 
Dictionnaire Encyclopddiq'ne LARO USSE. 
Larousse. 
\[Mil90\] Miller, G. et al. (1990). Fiw~ Papers on 
WordNet. CSL Report d2. Cog,fitive Science 
Laboratory, Princeton University. 
\[Oh1:78\] Ohtsuki, T. ct M. (1978). (hvw.n, l,~,'cnch,- 
Japanese Dictionary. Sanseido. 
\[Suz70\] Suzuki, S. (1970). Standard Jap,,,~,,se- 
Fre.nch Dictionaw. TMshuukan. 
\[Tmn85\] Tanmn~, T. (1985). Diction,mire F,',m~:aiSo 
gapo'nais R 0 YAL. H akusuislm. 
\[Tok90\] Tokmmga, T. mM Tanaka, 1\[. (1990). 
The Automat, it F, xtracdon of Concei)Lual 
Itmns from Bilingual l)icdonm'ie, s. I'IUCAI. 
Appendix 
The following two lemnu~s ~u'e needed to prow~ 
l'roi)erty 1. 
\]fminnia 1-1 :r G D.u..:,.:c e,=> y G D,_,,s:r 
This is dear t'l'Olll the symmeLric .qtFllCLlll'C or I.hc 
h;u'monizcd dicl,ion;try. 
Lc'ninia 1-2 If X 'is ~L sol, (every eleme,,.t ha.s 
a 'meigM of 1)then ~Sa(D ..... ~;X, Y) = 6a(X, Du-,:,'Y) 
Proof: (IX\] de.notes l~he mmfl)er of elemmits in 
x) 
5~(D:,.-.vX, :u) 
-= I{:,'k': ~ X A :U < D:, -.,:,,}1 
= \[{:,,l:r e X A :," G I)~, ...... Y}t 
= ~~(X,D v .... ,.q) t~ 
As branches do not ow~rhp, Df_.j is a set. From 
\]Jmuma l-2, the proof for Property 1 is giwm as 
follows: 
:=6a (Dr ,e:l:, DeJ,j j) 
==-D'a(Dj_~ef , Df +e f) \[\] 
'\])lie proof tln' t.lw s(~c(md equ:d:imt is I:he smne. 
303 
