A Method for Distinguishing Exceptional and G(m(.~L1 .... Examples-' ~ in 
Example-based Tr~msfer Systems 
Hideo W~tanal)c 
IBM th~search, Tokyo Research l,ahoratory 
1623-14, Shimotsuruma, Yamatt)-shi, Kanagawa-ken 242, JAPAN 
e-malh watanahe~)trl.vrmt.lbm.com 
Abstract 
Distinguishing exceptional translation examples is an 
important issue in example-based transflw systems, 
because such systems use exceptional and general 
translation examples unifi)rmly. This l);qmr ,lescribes 
a mechanism for dealing with exc,q)tiomd transla- 
tion examples in our example-hosed tnLnsfer system, 
Sim~lban, and proposes a method for identifying such 
examples in a translation exampl~base. 
1 Introduction 
and that we are given the folh)wing Japanese input 
s(,~.tence (sl): 
(sl) watashi(l)ha de,,t~ku(cah'ulator) w,, shiy- 
OIISH rll. 
In the almve exami,h~s , (s l) in likely to Im more sim- 
ilar to (,,1) than (e2), b,~c~use the three Japanese 
verbs "kyouyotmuru~" "tsakan," and "shiyousuru" 
are all w~ry simih~r, ~ and "dent~ku" ("calculator") 
is more similar to "konl)yuutaa" ("computer") than 
"kurum~'?' ("car"). If this is the case, the English out- 
put ohtained by using (el) is (tl),'-' whereas it should 
he 0,2): 
In recent years~ the example-based approach ha.s been 
used in many areas of natural language l)mcessing 
\[3, 7, 8, 10~ 9, 1\]. We haw~ been tlsing this al)- 
proach to develop a transfer system called 5'imTra', 
\[13, 14, 16\]. However, a bottleneck occured in the 
collection of large numbers of translathm examples 
consisting of pairs of parsed structures in the source 
and target languages (hereafter we (:all these struc- 
tures translation patterns)~ because parsing is not a 
perfi~et process. We now have some methods for over- 
coming this problem. For instance, recent studies 
\[2, 11, 6, 12\] have proposed mechanisms for collecting 
pairs of parsed structures automatically from transla- 
tion examples, and in the previous paper \[15\], 1 pro- 
posed ,~ method for extracting relevant translation 
patterns by comparing a wrong translation resttlL and 
its correct translation. Using these methods, we ca.n 
now collect translation patterns reh~tively easily. 
There is, however, another problem ca.lled e~:ample 
inte,¢e,~nce, wl,ich means that an ,xceptional (or id 
iomatie) translation pattern is selected when a gen- 
eral translation pattern should be selected; this has t~ 
side-effect on the construction of a target structure. 
Suppose that we have the following two translation 
examples from Japanese to English (el) and (e2), 
(el) watashi(1) ha konpyuuta~(computer) wo kyouy-- 
ollsllrll. 
I share the use of ~ computer. 
(e2) watashi(I) ha~ kurum~(ear) wo tsnkau. 
\]" /ISe a. car. 
(tl) I use the use of a calculator. 
(t2) I use a calculator. 
This probh~nl occurs because examph'~-ba,sed transfer 
systems choose examples simply on the basis of sim- 
ilarity. This ca.n be considered by using the analogy 
of cells like those shown in Figure 1. In the \[igure, a 
dot represents a translation e×aml)le ~ and a cell rep- 
resents a spac(! in which an input is determined to be 
similar. According to this analogy, all example-blused 
system chex:ks the cell in which an input is located, 
;~nd uses an ex~mple gow~rning the cell. If a new 
exa.mt)le is added in this space, it cell for it is cre- 
ated as if cell division. If an input happens to fall 
into the cell of an exceptional example, it is wrongly 
tr~mslated. Ther,d'ore~ an exce.ptitmal example shoukl 
be added as ~ spechd cell (a shaded dot in Figure 
1) that h~us no exte,t in the example-based space, so 
that it. cannot he used unless it matches the input ex- 
actly. Thus, an examl~le-based transfer system must 
deal with ,~xctq)tiomd translation patterns st:parately 
when calcuhtting similarity. 
This paper describes a mechanism used in Sim~lFan 
for dealing with exceptional translation patterns in 
the same framework as general translation patterns, 
and proposes a method for identifying exceptional 
tnrnsh~tion patterns in ~ tr~ulsla.tion pattern base. 
The next section describes a mechanism for dealing 
with such translatimL patter.s, and Section 3 de- 
1 Actually, they are in the same c~ttegory (or the same lmff) 
in the JItpttll~!Se thestmrus lh nr i-Go-l\[you \[5\]. 
r\['h\[~ Illltill Vftl'll iS /:llltllgl!¢l fl'tllll "shltrt!" to IU4t :" \]leCallSly 
"share" is not ~t transh~tiou of "shiyousuru." 
39 
Figure 1: An example-base space 
scribes a method for identifying exceptional trans- 
lation patterns. Some experiments are reported in 
Section 4, and some issues are discussed in Section 5. 
Finally, some concluding remarks bring this paper to 
an ond, 
2 Mechanism for dealing with 
exceptional translation pat- 
terns 
'kyouyou- 
suru' 
WO 
"kuruma" 
("car") 
...... "share" 
dobj 
"use" +the 
postmod 
"car" +of 
(tpl) 
"tsukau" " ........... 'use" lwo 
"kuruma" ............ "ear" 
(tp2) 
Figure 2: l';xeeptionM translation l)attern and general 
i, rans\[;tLion pattern 
SimTlNn calculates the similarity between a subgraph 
of an input structure and the source part of a transla- 
tion pattern on the basis of both the structural simi- 
larity and the similarity of the lexical-forms of cor- 
responding nodes. For instance, the distance (the 
inverse of similarity) between two Japanese lexical- 
forms is expressed by the difference of their values in 
a Japanese thesaurus called Bunrui-Goi-lIyou \[5\] 3 as 
follows: 
I@hcode(w,) - bgheode(~,,~)\[ + 
distance(wl, w2) = bghmax + b" 
where bghcode(w ) is the code vMue in the Bunrui- 
Goi-Hyou, bghma:c is the maximal difference of the 
bghcodes, and 6 is a penalty value incurred when wl 
and w2 are not identical. This equation is used for 
lexical-forms in general translation patterns. If one 
is a lexicM-form which requires exact-match in an ex- 
ceptional translation pattern, then the distance is cal- 
culated as follows: 
0 wl is identical to w 2 distance(wl, w2) 
1 othevwlse 
aBunrui-Goi-IIyou is a Japanese thesaurus consisting of 
large trees for nominals, adjectives, and verbs. Each node is as- 
signed a unique nmnber. Similar concept words are locattxl in 
similar positions (or assigned similar numbers) in these trees. 
A lexical-forni has a distinctive fea.tnre that makes it 
possible to determine which equation should be used 
hi cMculating similarity I if one of two le.xlcal-forms is 
expressed by a single-quoted string, then the distance 
between the lexical-forms is calculated by using the 
second equation; on the other hand, if both lexical- 
forms are expressed by double-quoted strings, then 
their distance is calc:nlated by using the first equation. 
Thus, an exceptional translation pattern is distin- 
guished by having nodes whose lexicM-forrns are 
single-quoted strings in its source part, while a 
general translation pattern is distinguished by hav- 
ing nodes whose lexicM-fi~rms are all double-quoted 
strings in its source part. Not MI nodes in the 
source part of an exceptional translation pattern are 
necessarily single-quoted strings; single-quoted string 
nodes and don bh+-quoted string nodes may be mixecl 
in a translation pattern, ht Figure 2, (tpl) is an ex- 
ceptional tr;ulslation pattern and (tp2) is a general 
translation pattern. Note tt~tt the root node of the 
Japanese part is the only single-quoted string in (tpl), 
and it matches only an input whose root node is 'ky- 
ouyoLIsHru. ~ 
By using this distinction of lexical-forrns, we e~n inte- 
grate exceptionality handling into the similarity cal- 
culation framework without separating this task as a 
pre process or post-process. ' 
40 
"kyouyou- " ..... "share" 
suru" i dob\] 
l wo "use" +the 
~ postmod 
"kuruma" • ....... ("ear") "car" +of 
(tpl) 
"tsukau" " ............ use" 1°°°, 
"kuruma" .............. car" 
(tp2) 
"tsukau" ........... "use" 
"\]itensya" "bicycle" 
(tp3) 
"tsukau" " ............ use" lwo 
"denwa ................ telephone" 
(in4) 
"lsukau" " ............ practice" 
I w° ,l d°b\] 
"mahou ................ magiC' 
(Ip5) 
Figure 3: F, xamI)lc ,)f the identilicati(nl 
3 Method for identifying ex- 
ceptional translation pat- 
terns 
For iriost peol)le , an exceptional translation pattern 
is likely to recall a pattern of translation for an i/l- 
iomatic or colloquial expression, hi generM, ;in id- 
iomatic translation pattern is a translation pattern 
whose target part is markedly different from that of 
translation patterns whose s¢)urce parts am similar to 
that of the idiomatic pattern. Froni the viewpoint of 
the transfer process, what we would like to identi\[y 
are translation patterns that may have side-effects 
when they are selected instead of general translation 
patterns. We call such translation patterns excep- 
tional travsIation pattern.s. According to this defi 
nltlon, exceptional translation patterns are not re- 
stricted to idiomatic patterns, in fact, more transla- 
tlon l)atterns other than idiomatic ones fall into this 
category. Here: we classify exceptional translation 
patterns into the following two categories: 
t Extra-Exceptional Translation l'atterns: These 
have some. extra elements hi the. target part in 
addition to those in similar traimhttimi patterns. 
i Intra-Exceptional "\]'rans\[atlon \]>atterns: These 
are almost same ms similar translation patterns, 
but several target words are different. 
of 0×eeptimial triulslatlon patterns 
When exceptional translation patterns are \[olind~ it is 
hnportant to know whether two translatiml patterns 
are e(lUivMent or not. '\]'herefore> equivalent transla- 
tion plctterns are defined as follows: 
(liven two dependency structures dl and d2, then 
they lore called equivalent if and only if tiiey are strlle- 
rurally identicM and correspmiding nodes have the 
similar seinantic code. 4 }"urther~ given two trails- 
lath,n patterns tp, = (si,ti,m,) tp2 = (s2,t2,m2), 
where .~i is ~L so)lr('e l)art, ti in a target part, an(I mi 
i~ a mapping from .~'i to iT, then these two transla- 
tion patterns ;ere called equivalent if they satisfy the 
following conditions: 
(1) Both sou roe parts axe equivalent> and both targ~t 
parts are strilctllrally identical. 
(~) 'l'he roots of l 1 a.nd 17 are the sallle strhlg. 
(3) For each ,m(le n hi .+~, ',n~(n)is o,,e of transhttion 
words of n. 
(4) t,'o~ each ,~o,le ,, in ,"2, ',,.,('n) is one or translation 
words of n. 
The. algorithm for identi\[yhlg e×ceptlonal trluislation 
patterns is as follows: 
,I \]?Of ili.~ltltlll It~ tiil~ :'ll!lllitllt, iC code ill JIL|llllll!~\[~ \[~+ ~llllrll\[-(loi- 
Hyou code. The extent to whh:h two words are determhw, d to 
I,c similar is *also a p~ranleter. It may vary according to the 
system. In this liltper, two words iu't~ deternllncd to be similar 
if they have the ~anle senuu~tic c,Me. 
4"/ 
Step 1 Divide translation patterns into sew~ral groups, 
each of which consists of equlwdent translation 
patterns. 
Step 2 For each pair of distinct translation pattern 
groups gl and g~, if any pattern of 9t is equiva- 
lent to any pattern of g2 other than nodes gov- 
erned by the root of the source l)art, tlmn the 
translation patterns in gl arid 92 are marked gen- 
er'~L 
Step 3 ~br each pair of distinct translation pattern 
groups gx and g2, if" the source part of any pat- 
tern (pl) of gl is equivalent to the source part of 
any pattern of g2, but target parts of them are 
not struetnrally identical, because Pl ha.s extra 
elements~ then the translation patterns of gl are 
marked extm-exeeptionaL 
Step 4 For each non-exceptional translation pattern 
group gl, if there is another general translation 
pattern group g~ such that any pattern (Pl) of 
gl is equivMent to any pattern of g2 other than 
the root node in the target part of Pt, then 
the translation patterns of gt are marked itth'a- 
exceptional. 
Step 2 identifies possible general translation patterns 
if they are used in a relatively wide range of'words, be- 
cause in general an exceptional pattern is restricted 
in the usage of words. This approach, however, is 
not perfect rot identif,ying general translation pat- 
terns, becanse there in ~t c~use such that the exccp- 
tionality derives from a single special word. There- 
fore, in the next step, checking does riot exclude these 
possible general translation patterns. Step 3 identi- 
ties extra-exceptional translation patterns by check- 
ing the structure of the target part. Step 4 then iden- 
tifies intra-exceptional ones by comparing the mot 
node in the target part with the root nodes in the tar- 
get part of possible general translation patterns. The 
reason why this comparison is restricted to possible 
general translation patterns is that intra-excepti(n,d 
translation patterns have si(h~efrects only when they 
are similar to general translation patterns. 
Figure 3 shows an example of the identiflcation 
of exceptional translation patterns, in which the 
Japanese verbs "kyouyousuru" and "tsukau" haw.' the 
same bghcode, and the Japanese nouns "kuruma," 
"denwa~" and "mahou" have different bghcodes, on 
the other hand, "kuruma" and "jitensyd' have the 
same bg|,eode. First, step 1 divides tImse transla- 
tion patterns into four groups: group 1 c.onsists of 
(tpl), group 2 consists of (tp2) and (tp3), group 3 
consists of (tp4), and group 4 consists of (tp5). Step 
2 identifies group 2 and 3 as general translation pat- 
terns, because "kuruma" and "denwa" have different 
bghcodes. Subsequently, step 3 identifies (tpl) as an 
extra-exceptional translation pattern, beci~use (tpl) 
has extra elements "the use of" for (tp~). Further, 
step 4 identifies (tpS) as at, iutra-exceptional transla- 
tion pattern, because (tp5) is equivalent to the gen- 
eral translation patterns (tp2), (tp3) and (tp4), other 
than "use" and "practice" in the root nodes of the 
target parts. 
4 Experiments 
We have tested the almve-nientioned algorithm with 
translation patterns in a Japanese-to-English trans- 
fer dictionary that was previously used in our lab- 
oratory. For each bghcode, we. collected translation 
patterns such that the root of the source part has 
the. code. and a.pplied the algorithm to tim transla- 
tion pattern set of each category. Table 1 shows the 
resulting top 10 categories with respect to tt,e total 
number of occurrences. In most categorles, more than 
90% of translation patterns were identified as excep- 
tional. The reason for the lopsidedness of, this result 
is that tl,e translation patterns described in the pr(~ 
vious transfer dictionary were almost all exceptional 
eases that conhl not be. de.all with by the default pro- 
cedures coded in the transfer module. Therefore, this 
result indicates that the ~dgorithm is able to idenitfy 
exceptional translation patterns correctly. 
5 Discussion 
In conventhmal tra, nsfl,~r systems \[4\], transfer rules are 
roughly divhled into general ones and exceptional (or 
idiomatic) ones. The transfer system checks the ex- 
cepth)nal ca.ses first, and if they cannot match the 
input then the system applies general rules. On the 
other hand, example-based transfer systems deal with 
translation patterns (or examples) uniformly on the 
basis of similarity, according to the example-b~sed 
pri,ciph,. 'rids m~ci,mism causes the exanlple in- 
terference problem. A very useful property of the 
e×ample-I)~u~e(l approach is that it allows a sente.nce 
to be added as an examph~ if it cannot be dealt with 
properly. This holds if the same input :~s the newly 
added example is given~ but when the resolution of 
the slmilarity calculation is not enough, an input that 
is similar to but not exactly the same as the added ex- 
ample may not be dealt with properly, because there 
may be another similar example that is exceptional. 
'l'hereh)re, it is very important to identify whether an 
example is general or exceptional. 
After application of the alg<>rithm described in this 
paper, translation patterns are classified into the fol-. 
lowing categories: general, exceptional (extra- and 
intra-), and neutral. Neutlal translation patterns, 
which are not ml~rke.d general or exceptional, are 
42 
Bghcode 
(example) 
15210(idousuru) 
15270(iku) 
15310(torikomu) 
15600(tikazuku) 
15710(kiru) 
30110(kurushimu) 
30200(suki) 
30610(mnou) 
31200(iu) 
36700(hattyuu) 
38520(tsukau) 
Num of \] Num of-\[ Num of l';xceptio,ml 
( xtra, i,,tra) 
247 \] 1 232 (228, 4) 
174 \] 0 138 (137, 1) 
365 I 0 160 (150, 10) 
199 , 1 185 (178, 7) 
185 0 181 (159, 22) 
192 8 183 (160, 23) 
280 6 271 (203, 68) 
180 0 179 (169, 10) 
191 0 173 (17a, 0) 
18'2 0 181 (108, 13) 
65 ~ 00 (53, 7) 
ExeeI,tional (extra n,,l~,) 
/Total 
93% (9~%) 
79% (78%) 
96% (90%) 
92% (89%) 
97% (85%) 
95% (83%) 
96% (72%) 
99% (93%) 
90% (90%) 
99% (92%) 
!m% (81%) 
Tabh.' 1: Experime,ltal results for transfer dictionary 
translation patterns that do not h~ve sld~>effects. 
They are n(~t used for a wide variety of words in 
the current translation p~tttern bmse. If m~)re trans- 
lation patterns are added later, they m~ty be identi- 
fied as general or exceptional. By this method, mm 
can enable the system to identify exceptional trans- 
lation patterns automatically hy adding some general 
translation I>atterns similar to them. This is a very 
useful feature for bootstrapping of ~t transh~ti<m pat- 
tern base. A weak point of this algorithm, }mwever, 
is that it requires a large number of translation pat- 
terns. If enough translatiml patterns ;~re not given, 
exceptional translation l)atterns might n(,t be identi 
tie(\[, tlowever, collecting many tr;ulslatinn patterns 
is no longer a serious l)roblern, since several methods 
for eolleeti,ig them automatically have been pr/q)(Ised 
in recent studies \[2, 11, 14, 6\]. 
The method proposed in this paper probad)ly does 
not comply with human intuition regarding idiomatic 
translation patterns; rather, it detects transh~timt 
patterns that are idiomatic for the system, in other 
words, patterns that might have side-effects in the 
current set of translati(m patte.rns. It prnl,ahly re- 
quires deeper scm~mtle pr()cessing to ide.nti fy transht- 
tion patterns tlu~t are idiomt~tie in the conventional 
Sellse. 
6 Conclusion 
In this paper, we have showll a problem of example- 
based transfer systems, example inlerfl'lv*u:e, and de- 
scribed a mechanism for dealing with exceptional 
translation patterns and general translatitm p;ttterns 
uniformly in similarity calculation withmlt destroy- 
ins the whole framework of example-bmsed process- 
ing. Further, we have proposed a method fi)r disl.in- 
guishing exceptional translation patterns from gen- 
eral translation patterns. In some cases, this met.h<nt 
giw~s results that do not match human intuition re- 
garding idiomatic translation patterns, but it can de- 
tect, from the viewpoint of example-based processing, 
tra.nslatiml p~ttterns in the current translation pattern 
base that might have side-effects. 
References 
\[1\] Furuse, O. and Iida, 1l., "Cooperation between Trans- 
fer and Analysis in l"~xampleqb~sed Framewc~rk," Proc. 
of Coling 92, Vol. 2, pp. 648-681, 1992. 
\[2\] Kaji, 1\[., Kid% Y., and Morimoto, Y., "Learning 
q'r~mshttion 't'omphttes from Bilingual Text," f'voe, of 
Colin.q 92, Vol. 2, pp. 672 678, 1992 
\[31 Nagao, M., "A l,'ramework of ~t Mechanical 'l'ransla- 
tion between Japanese ~u,d English by Analogy Prin- 
ciple," I;'Jitlmrn, A. and }lanerji, 1{. (eds.) : Artificial 
and Human lr~.lelhgence, NA'I'O 1984. 
\[4\] Nagao, M., "The Transfer Phase of the Mu Machine 
'\['ranslation System," l'roe, of Coling '80 pp. 97--103, 
1986. 
\[5\] Nal.iomtl l,angmtge I{.esearch lnstil.ute: lhmrui Goi 
llyou (in Japanese), Syuuei Syuppan, 1964. 
\[\[3\] Matsumato, Y., Ishinmt.o, II., and IJtsuro, T., "Struc- 
I.ural Mal.ching or Par/tllel Text," l'roc, of 31st Annual 
Meelin9 of ACL, pp. 23 30, 1993. 
\[7\] Sadler, V., "Working with Analogical Semantics," 
l,'oris Ih,blicat.ions, 1989. 
\[8\] Sata, S. and Nag;to, M., "Toward Memory-based 
'l'ranshtl.ion," Col*ng 90, \]990. 
\[9\] Sat{), S., "Memory-tntsed Translation," Doctor 7'hesis, 
1992. 
\[10\] Smnita, E., lids, 11., and Kohyama, I1., "q¥anslating 
with Examples: A New Approach to Machine q¥ans- 
lotion," l'roc, of lnfo aap,n 90, 1990. 
\[11\] Utsuro, T., Matsumoto, Y., and Nagaa, M., "l,ex- 
ical Kn¢>wledge Acquisitiofi from Bilingual Corporlt," 
l)rac, of C¢)ling '92, Vol. 2,'pp. 581-587, 1992. 
43 
\[12\] Utsuro, T., Matsumoto, Y., and Nagao, M., "Ver- 
bal Case Frame Acquisition from flilingual Corpora," 
Proc. of IJCAI '93, Vol. 2, pp. 1150-1156, 1993. 
\[13\] Watanabe, tt., "A Model of a qYansfer Process Using 
Combinations ofTranslatlon ltules," Proe. of Pacific 
Rim of Inl. Conf. on AI '90, 1990. 
\[14\] Watanabe, 1I., "A Simitarlty-Driven rl'ransfer Sys- 
tem," Proc. of Coling '92, Vol. 2, pp. 770 -776, 1992. 
\[15\] Watanabe, ti., "A Method for F, xtracting rPransla- 
tion Patterns from '\[¥anslation F, xamples," Proc. of 
5lh Int. Conf. on Theoretical and Methodological Is- 
sues in Machine 7~'anslalion, pp. 292-301, 1993. 
\[16\] Watanabe, tI. and Maruyama, If., "A qYansfer Sys- 
tem Using Example-Based Approach," IEICIi'~ '\['tans- 
actions on Information and Systems, Vol. E77-1), No. 
2, pp. 247-257, Feb. 1994. 
44 
