T ansla ion Memo y s. Example-based MT –
Wha ’s he di e ence?
HAROLD SOMERS (UMIST, Manches e )
GABRIELA FERNANDEZ DIAZ (Uni e sidad de Se illa)
1 In oduc ion
Th oughou he 1970s and 1980s, Machine T ansla ion (MT) esea ch ocused on he
de elopmen o so-called “second gene a ion” sys ems, which aimed o ansla e ex
by a p ocess o ule-d i en linguis ic p ocessing, usually in h ee s ages: syn ac ic-
seman ic analysis o he sou ce ex , bilingual ans e a a mo e o less abs ac le el
o ep esen a ion, and a ge - ex gene a ion om syn ac ic ep esen a ion. A he
same ime, om a p ac ical poin o iew, he e was much discussion on how sys ems
buil wi h his a chi ec u e could be used o deli e a easonable s anda d o
ansla ion usable by eal use s. Mos popula we e ideas o es ic ing he inpu ( he
sublanguage and con olled language app oaches), o in ol ing he use in p e- and
pos -edi ing. In e ac i e MT, whe e he use and he compu e would coope a e in
esol ing ambigui ies and making choices, was also championed.
In he ea ly 1990s, wi h hese ideas ai ly well es ablished, and pe haps e en
g owing s ale, esea ch in MT was hi by an appa en ly new pa adigm in which in
pa icula he eliance on linguis ic ule sys ems was o be (a leas pa ially) eplaced
wi h he use o a co pus o al eady- ansla ed examples which would se e as models
o he MT sys em on which o base i s new ansla ion. This app oach came o be
known as Example-Based MT (EBMT) which had in ac i s been p oposed en
yea s ea lie , in 1981. We will look in a li le mo e de ail a his de elopmen below.
A abou he same ime, a new ool o ansla o s was being men ioned by
de elope s. Like EBMT, i used a co pus o al eady- ansla ed examples o se e as
models o he new ansla ion, bu c ucially, i was he human use s, no he compu e
i sel , who should de e mine exac ly how o use he examples in p oducing a new
ansla ion. This ool is o cou se now widely known as a T ansla ion Memo y Sys em
(TMS).
1
The e a e many commen a o s who ega d EBMT and TMSs as essen ially he
same hing, and indeed ce ain de elopmen s in bo h ha e b ough hem close o each
o he . The p esen au ho has long main ained ha he e is a c ucial di e ence, which
will be elabo a ed in he nex sec ions. Howe e , he e a e simila i ies, and in
pa icula p oposals o make TMSs be e in a iably make hem look mo e and mo e
like EBMT sys ems. This, p incipally, is he subjec o his pape .
1
We will h oughou dis inguish he sys em om i s p incipal componen , he da abase o s o ed
ansla ions, i.e. he “memo y”. Fo his eason we will e e o he sys ems as TMSs, and o he
da abase i sel as a TM.
2 In e wined his o y o TMSs and EBMT
A a isk o epea ing poin s ha a e amilia o eade s o his jou nal, we wish in his
sec ion o desc ibe some key momen s in he de elopmen o he concep o TM(S)s.
The o iginal idea o a TM is usually a ibu ed o Ma in Kay’s well-known “P ope
Place” pape (1980), al hough he de ails a e only hin ed a obliquely:
[T]he ansla o migh s a by issuing a command causing he sys em o display any hing in he
s o e ha migh be ele an o [ he ex o be ansla ed] .... Be o e going on, he can examine
pas and u u e agmen s o ex ha con ain simila ma e ial. (Kay, 1980:19)
In e es ingly, Kay was pessimis ic abou any o his ideas o wha he called a
“T ansla o ’s Amanuensis” e e ac ually being implemen ed. Bu Kay’s obse a ions
a e ac ually p eda ed by he sugges ion by Pe e A he n (1978)
2
ha ansla o s can
bene i om on-line access o simila , al eady ansla ed documen s, and in a ollow-
up a icle, A he n’s p oposals qui e clea ly desc ibe wha we now call TM(S)s:
I mus in ac be possible o p oduce a p og amme [sic] which would enable he wo d p ocesso
o ‘ emembe ’ whe he any pa o a new ex yped in o i had al eady been ansla ed, and o
e ch his pa , oge he wi h he ansla ion which had al eady been ansla ed, ....
Any new ex would be yped in o a wo d p ocessing s a ion, and as i was being yped, he
sys em would check his ex agains he ea lie ex s s o ed in i s memo y, oge he wi h i s
ansla ion in o all he o he o icial languages [o he Eu opean Communi y]. ... One ad an age
o e machine ansla ion p ope would be ha all he passages so e ie ed would be
g amma ically co ec . In e ec , we should be ope a ing an elec onic ‘cu and s ick’ p ocess
which would, acco ding o my calcula ions, sa e a leas 15 pe cen o he ime which
ansla o s now employ in e ec i ely p oducing ansla ions. (A he n, 1981:318).
Alan Melby (1995:225 ) sugges s ha he idea migh ha e o igina ed wi h his
g oup a B igham Young Uni e si y (BYU) in he 1970s. Wha is ce ain is ha he
idea was inco po a ed, in a e y limi ed way, om abou 1981 in
ALPS
, one o he i s
comme cially a ailable MT sys ems, de eloped by pe sonnel om BYU. This ool
was called “Repe i ions P ocessing”, and was limi ed o inding exac ma ches modulo
alphanume ic s ings.
3
The much mo e in en i e name o “ ansla ion memo y” does
no seem o ha e come in o use un il much la e .
4
The i s TMSs ha we e ac ually implemen ed, apa om he la gely in lexible
ALPS
ool, appea o ha e been Sumi a and Tsu sumi’s (1988)
ETOC
(“
E
asy
TO
C
onsul ”), and Sadle and Vendelman’s (1990) Bilingual Knowledge Bank, p eda ing
wo k on co pus alignmen which, acco ding o Hu chins (1998) was he p e equisi e
o e ec i e implemen a ions o he TM idea. Also, Kugle e al. (1991) epo on
wo k by Keck (1989) based on s a is ical me hods in he con ex o an
ESPRIT
esea ch
p ojec . I is di icul o pinpoin when TMs en e ed he consciousness o ansla ion
s udies esea che s and ansla o s in gene al. B ian Ha is, in oducing he no ion o a
“bi- ex ” in a ansla o s’ magazine, p oposes some hing like a TM wi hou using ha
2
Ea ly p oposals o a TM, and o he aspec s o he idea o a T ansla o ’s Wo ks a ion a e desc ibed in
Hu chins (1998).
3
Cu iously, his mos inno a i e ea u e is ba ely men ioned in desc ip ions o he ALPS sys em, wo
excep ions being Sibley (1988:96,100) and Wea e (1988:121 ) who men ion he acili y almos as an
a e hough .
4
Ex ensi e enqui ies ha e so a ailed o p oduce a sa is ac o y iden i ica ion o he i s use o his
e m. Hu chins (1998:303) sugges s ha he T ados company we e he i s o use he e m.
name (Ha is, 1988:9):
5
a da abase o pai ed ansla ions, sea chable ei he by
indi idual wo d, o by “a whole ansla ion uni ”, in he la e case he sea ch being
allowed o e ie e simila a he han iden ical uni s. Ca e (1988) esponded o ha
a icle wi h an announcemen ha Logos we e ma ke ing jus such a ool. In an
unsigned 1991 a icle he magazine Language In e na ional epo ed ha “ ex banks”
had now made hei appea ance:
The i s o hese would seem o ha e been IBM Eu opean Language Se ice’s T ansla ion
Suppo Facili y (TSF), which, …, inco po a ed a epea ed sen ence iden i ica ion acili y.
(Anon, 1991:5; emphasis added).
The a icle goes on o explain he no ion o “ uzzyma ch” (sic) in he case whe e exac
ma ches a e no ound.
The p oceedings o Aslib’s indica i e annual con e ence se ies T ansla ing and
he Compu e con ain no men ion a all o TMs un il 1992, when h ee sepa a e
a icles (F eibo , 1992; Le-Hong e al., 1992; S anholm, 1992) men ion hem, in one
case (Le-Hong e al.) wi hou eeling he need o explain he e m. B ace (1992)
epo ed de elopmen o a TM ool by T ados, as well as he
ESPRIT
p ojec men ioned
abo e, and p ojec s a IBM’s Eu opean Language Se ices (Denma k) and he
O icial Languages and T ansla ion sec o o he Canadian Depa men o he
Sec e a y o S a e in O awa.
The idea o EBMT has a simila ch onology, wi h ideas su acing in he ea ly
1980s ( he pape p esen ed by Mako o Nagao a a 1981 con e ence was no published
un il h ee yea s la e – Nagao, 1984), bu he main de elopmen s being epo ed om
abou 1990 onwa ds.
6
The essence o EBMT, called “machine ansla ion by example-
guided in e ence, o machine ansla ion by he analogy p inciple” by Nagao, is
succinc ly cap u ed by his much quo ed s a emen :
Man does no ansla e a simple sen ence by doing deep linguis ic analysis, a he , Man does
ansla ion, i s , by p ope ly decomposing an inpu sen ence in o ce ain agmen al ph ases ...,
hen by ansla ing hese ph ases in o o he language ph ases, and inally by p ope ly composing
hese agmen al ansla ions in o one long sen ence. The ansla ion o each agmen al ph ase
will be done by he analogy ansla ion p inciple wi h p ope examples as i s e e ence. (Nagao,
1984:178 )
Nagao co ec ly iden i ied he h ee main componen s o EBMT: ma ching agmen s
agains a da abase o eal examples, iden i ying he co esponding ansla ion
agmen s, and hen ecombining hese o gi e he a ge ex . Clea ly EBMT in ol es
wo impo an and di icul s eps beyond he ma ching ask which i sha es wi h TMS.
The idea o EBMT eally ook o in he ea ly 1990s, wi h an inc easing numbe
o pape s a con e ences epo ing on his app oach. Pionee s we e mainly in Japan,
including Sa o and Nagao (1990) and Sumi a e al. (1990). Men ion should also be
made o he wo k o he DLT g oup in U ech , o en igno ed in discussions o
EBMT, bu da ing om abou he same ime as (and p obably wi hou knowledge o )
Nagao’s wo k. The ma ching echnique sugges ed by Nagao in ol es measu ing he
seman ic p oximi y o he wo ds, using a hesau us. A simila idea is ound in DLT’s
“Linguis ic Knowledge Bank” o example ph ases desc ibed in Pappegaaij e al.
5
In he nex pa ag aph, he desc ibes i as p o iding “a memo y-pe ec exploi a ion o he ansla o ’s
own p e ious expe ience”.
6
A ho ough e iew o he li e a u e on EBMT is a emp ed in Some s (1999).
(1986a,b) and Schube (1986:137 ). Sadle ’s (1991) “Bilingual Knowledge Bank”
clea ly lies wi hin he EBMT pa adigm.
Du ing his ea ly pe iod, indi idual esea che s in he ield used al e na i e
names, pe haps wan ing o b ing ou some key di e ence ha dis inguished hei own
app oach: “case-based” (Collins and Cunningham, 1996), “analogy-based” (Nagao,
1984), and “expe ience-guided” (Zhao and Tsujii, 1999) a e all e ms ha ha e been
used. The i s o hese ecalls app oaches o Machine Lea ning known as “case-based
easoning” (Riesbeck and Schank, 1989), and o he ela ed models.
7
Ano he e m
ound is “memo y-based ansla ion” (Sa o and Nagao, 1990; Ki ano, 1993), he use
o which p obably did mos o sugges a ini ies be ween EBMT and TMSs.
3 Wha EBMT and TMSs could ha e in common
EBMT and TMSs ha e in common he use o a da abase o p e ious ansla ions, he
“memo y” o “example-base”, and he essen ial i s s ep, gi en a piece o ex o
ansla e, o inding in he example da abase he bes ma ch(es) o ha ex . Once he
ma ch has been ound, he wo echniques begin o di e ge. Howe e , i would be
misleading o assume ha all hey ha e in common is he ask o ma ching, o e en
ha he app oaches o ma ching in he wo camps a e pa icula ly simila . Use o a
da abase implies issues o da abase design, con en , and main enance. These will be
he ocus o he nex sec ions.
3.1 How a e examples ound?
In TMSs, he TM da abase i sel can be cons uc ed in one o h ee ways. The
simples me hod, hough he mos ime-consuming one, called “in e ac i e
ansla ion” by Bowke (2002:108 ) is o build a TM om sc a ch, ha is, o s o e in
he memo y each sen ence as i is ansla ed. A second me hod, e e ed o by Bowke
(2002:109 ) as “pos - ansla ion alignmen ”, and much he alded by manu ac u e s, is
o ex ac a TM om an al eady ansla ed ex by aligning he sou ce and a ge ex s.
This can be a mo e o less i ksome ask (c . Macdonald, 2001), and he e is a
conside able li e a u e desc ibing a ious alignmen me hods in ol ing di e ing
amoun s o (linguis ic) sophis ica ion (see Manning and Schü ze, 1999:466–486 o
Wu, 2000a). O’B ien concludes:
A ansla ion memo y is always mo e accu a e when i has been c ea ed by in e ac i e
ansla ion as opposed o au oma ic alignmen , bu alignmen can p oduce a easonably accu a e
ansla ion memo y which can be used as a s a -up. (O’B ien, 1998:119)
Finally, TMs ha ha e al eady been c ea ed can be impo ed, and he es ablishmen o
ag eed in e change o ma s be ween manu ac u e s has hugely acili a ed his (no ably
he T ansla ion Memo y eXchange (TMX) o ma de eloped by LISA (Localisa ion
Indus y S anda ds Associa ion, c . Melby, 1998, 2000 and Topping, 2000).
The size o he TM is an ob ious ques ion. The TM li e a u e says li le mo e
han “ he bigge he be e ”, subjec o p ocessing limi a ions, hough Bowke
(2002:108) wa ns ha “size should no come a he expense o o ganiza ion”,
sugges ing ha sepa a e TMs o di e en subjec ields o clien s may help o educe
alse hi s caused by homonymy. She adds:
7
The ela ionship be ween EBMT and Case-Based easoning is discussed in Some s and Collins (in
p ess).
Keep in mind ha a la ge TM will esul in a g ea e numbe o ma ches .... The e o e, while i
may seem logical a i s glance o build a single la ge TM ..., his may u n ou o be a alse
economy.... Mo eo e , he e is a g ea e likelihood o e ie ing “noise” (e.g., ma ches ha a e
no help ul, ma ches con aining homonyms) and he ansla o may was e a conside able amoun
o ime analyzing, elimina ing, o edi ing hese poo ma ches. (idem.)
Fo Heyn (1998), a “big” TM will ha e be ween 100,000 and 1 million uni s, hanks
o ecen echnology ad ances.
In ecen spa sely-coded-ma ix based sys ems, eal in e ac i e wo k on ‘big’ mas e ansla ion
memo ies is possible. Big ansla ion memo ies a e ypically in he o de o 100,000 ansla ion
uni s, al hough memo ies in he ange o 500,000 o 1,000,000 ansla ion uni s a e en isaged by
he end o 1997. Acco ding o cu en esea ch es ima es, ansla ion memo ies could be made
up o 40% bigge wi hou any inc ease in cons an access imes. (Heyn, 1998:128)
In he EBMT li e a u e, he sizes o he example-bases epo ed a y o e a
huge ange, wi h 0.73m he bigges , and 7 (se en!) he smalles epo ed (c . Some s,
1999:120). Ob iously, he sys ems wi h iny example-bases a e pu ely expe imen al,
while he mo e se ious sys ems will ha e housands o examples ( a he han, say,
hund eds).
Use manuals o TMSs sugges e ising he da abase e e y so o en o clea ou
useless examples. By “useless” is p esumably mean “no used” a he han, o
example, “misleading” (c . Heyn, 1998:131 ). I is easy o see how a TMS could
inco po a e a measu e o he o me , simply by coun ing access. To measu e he la e ,
i would also need o “know” wha he ansla o is doing wi h he p oposed ma ch.
The “sui abili y” o examples is add essed in he con ex o EBMT sys ems by a ious
esea che s. Nomiyama (1992) in oduces he no ion o “excep ional examples”, an
idea u he de eloped by Wa anabe (1994). As a as can be seen, hese examples a e
excep ional jus in he sense ha i used hey gi e he w ong esul ! Clea ly a mo e
sys ema ic no ion is needed. I is well known ha he same ph ase can be ansla ed
di e en ly in di e en ci cums ances. Ellipsis, anapho a and s ylis ic a ia ion can
con ibu e o his, in which case di e en examples may be seen as none heless
equi alen in some sense. On he o he hand, he unde lying meaning o a ph ase may
di e depending on he con ex . Some s e al. (1990:274) illus a e how he simple
ph ase OK in a con e sa ion may be ansla ed in o Japanese as waka imashi a ‘I
unde s and’, iidesu yo ‘I ag ee’ o ijō desu ‘le ’s change he subjec ’.
The e is also an issue o “g anula i y”: bo h in TM and EBMT, he e is a ade-
o be ween leng h and simila i y o examples. The longe he example uni s, he
lowe he chance o an exac ma ch; bu he sho e he uni s, he g ea e he
p obabili y o “ambigui y” (mul iple, con lic ing, ma ches), wi h a co esponding
dec ease in he quali y o he p oposed ansla ion. Ni enbu g e al. (1993:48) call his
“passage bounda y ic ion and inco ec chunking”. The ob ious and in ui i e “g ain
size” o examples, o judge om almos all TMSs and EBMT sys ems, is he
sen ence, hough e idence om ansla ion s udies (Ge lo , 1987; McTai e al.,
1999) sugges s o he wise: human ansla o s p ocess ex in “na u ally-occu ing
syn ac ic uni s” and “gene ally he e is e y li le p ocessing a sen ence le el”
(McTai e al., 1999). Acco ding o Benne ,
... he e a e good easons o keeping he U[ni o ] T[ ansla ion] (in he sense o ansla ion
a om) in MT as small – and hence as manageable – as possible. Adop ing a la ge UT may be
less e icien , and is no gua an eed o imp o e ansla ion quali y. (Benne , 1994:18)
he “ ansla ion a om” being he smalles segmen ha mus be ansla ed as a whole
(ibid., p.13). Schäle e al. (2003) echo his sen imen , sugges ing ha “ma ching

segmen s a sen ence le el unnecessa ily es ic s he po en ial and he use ulness o
ansla ion memo ies” (p. 89), and p opose “ph asal ma ching” as he p ima y
mechanism o TMSs. Sima d (2003) s udied how eal use s make use o a bilingual
conco dance , a ool which closely esembles a TMS in unc ion, excep ha use s can
look up a bi a y sequences o wo ds. He ound ha mos use s look up syn ac ically
well- o med “chunks”, and implemen ed a sys em based on his p inciple. In ac ,
ma ching segmen s a he han whole sen ences p oduces a oo many “hi s”, so he
sys em mus also ha e a way o selec ing he mos use ul om amongs hem. An
e alua ion o Sima d’s implemen a ion sugges ed ha i p oposed be ween 15 and 30
imes mo e “ eusable ma e ial” han a sen ence-based sys em.
Bo h EBMT and TMSs could p obably be imp o ed by concen a ing on a mo e
lexible iew o he uni o ma ching/ ansla ion, and “ he exploi a ion o agmen s o
ex smalle han sen ences” (C anias e al., 1994:100). In TMSs, his idea is pa ly
add essed in ha e minology look-up is o en seen as an in eg al pa o he ool,
al hough e minology ools a e gene ally implemen ed in a lexicon-based a he han a
memo y-based manne . We will e u n o he issue o “ agmen s” below. Acco ding
o Bowke ,
Many TM sys ems allow he use o de ine o he uni s o segmen a ion in addi ion o sen ences.
These uni s can include sen ence agmen s o e en en i e pa ag aphs. (Bowke , 2002:94)
Along he same lines, Esselink s a es:
A segmen is a ex elemen , which is conside ed by he applica ion as he smalles ansla able
uni , de ined by pe iods, semi-colons, and ha d e u ns. These a e usually sen ences, bu can also
be chap e headings o i ems in a lis . ....T ansla ion memo y ools usually allow he use o
change and cus omize segmen a ion ules. (Esselink, 2000:362 )
3.2 How a e examples s o ed?
In TMSs, examples a e gene ally s o ed as plain ex , some imes wi h o ma ing
in o ma ion. Sys ems di e as o how hey ea o ma ing (i.e. on s, capi aliza ion
and so on) e en hough i is po en ially e y use ul o ma ching (see below).
Aus e mühl commen s ha
Some ansla ion memo ies ha e a buil -in in e ace ha wo ks wi h common wo d-p ocesso s
... he o ma in which he ansla ed ex is s o ed in he ansla ion memo y is iden ical o ha
used in he wo d-p ocessing p og am. (Aus e mühl, 2002:138)
F om his we mus in e ha segmen s always keep hei o iginal o ma when s o ed.
A di e en issue is he way TMS deal wi h ags. Nowadays, ansla ion ools a e
being used ex ensi ely in he so wa e localiza ion indus y. Fo his eason, he new
gene a ion o TMSs such as T ados, T ansi and Déjà Vu con ain a wide ange o
il e s o con e iles om one o ma o ano he . A he same ime, TMSs a e
designed o handle a wide a ie y o o ma s such as HTML, SGML and XML.
Esselink con i med he ollowing in he yea 2000:
Mos ansla ion memo y ools ha e s anda d il e s o HTML iles. HTML iles usually
con ain e y epe i i e ex , so i is wo hwhile using ansla ion memo y, because o he
subs an ial ime and cos sa ings. Fu he mo e, ansla ing upda es o web si es is much easie
and quicke i a ansla ion memo y o he p e ious e sion exis s.
Examples o ansla ion memo y ools ha suppo he HTML o ma a e T ados T ansla o ’s
Wo kbench, IBM T ansla ionManage , STAR T ansi , SDLX, and A il Déjà Vu. (Esselink,
2000:218)
In 2003 new e sions o TMSs ha e sp ead h oughou he ma ke . These new
e sions a e well equipped o deal wi h e e y kind o applica ion o he design o
web pages, p esen a ions, g aphics, e c. Thus, T ansi XV includes a numbe o il e s
ha make i possible o ansla e iles gene a ed wi h p og ams such as Excel,
Powe Poin , Qua kXP ess, PageMake , F on Make and Au oCAD, among o he s. In
he same way, he la es solu ions p esen ed in he ma ke by T ados and Déjà Vu,
namely, T ados 6.5 and Déjà Vu X, inco po a e il e s which allow he use o impo
and expo iles wi h any kind o o ma .
Inc easingly, TMS de elope s a e ecognising he alue o inco po a ing “ma k-
up” in o hei sys ems, no jus o ma ing bu also linguis ic anno a ions such as
syn ac ic pa -o -speech (POS) ags. In his espec , Planas (1999:8), s a es ha he
Xe ox XMS Memo y Manage , is a “linguis ically based ool”, and consequen ly is
capable o e ie ing be e ma ches han cha ac e -based sys ems. We ead “ his
shows he c ucial impo ance o using linguis ic da a o enabling mo e p ecise
e ie al o he closes sen ence in he da abase”. This sys em is cu en ly s ill in he
expe imen al s age howe e . Planas and Fu use (1999) p oposed a much mo e
elabo a e scheme in which examples a e ep esen ed in a mul i-le el la ice,
combining ypog aphic, o hog aphical, lexical, syn ac ic and o he in o ma ion. A
majo d awback o he mos success ul TMSs a ailable in he ma ke has o do wi h
he lack o inco po a ing linguis ic knowledge in hei p oduc s.
Ob iously, s o age and ma ching me hods a e in ica ely ela ed: we will e u n
o he la e in he nex sec ion.
In EBMT sys ems, a wide ange o o ma s ha e been p oposed o s o ing he
examples. Gi en i s o igins as a a ian o ule-based MT, ea ly EBMT sys ems
supposed ha examples would be s o ed as aligned ee s uc u es such as he one
illus a ed in Figu e 1, om Wa anabe (1992).
Figu e 1. Rep esen a ion schema o Kanojo wa kami ga nagai (li . ‘she
TOPIC
hai
SUBJ
is-long’)
↔
She has long hai .
Un o una ely, such a ep esen a ion in ol es se ious o e heads in s o age space,
analysis a un ime, and e i ica ion o s uc u es, a c i icism ha also applies o
Planas and Fu use’s p oposal, as shown in Figu e 2. The esul ing eliance on pa sing
o o he knowledge- ich p ocesses is acknowledged as a disad an age.
kanojo
nagai
kami
wa ga
ha e
she hai
long
subj obj
mod
Because hese ich ep esen a ions a e widesp ead in ea ly EBMT p oposals,
hey a e o en hough o as being a necessa y ea u e o EBMT, hough his is qui e
inco ec . La e EBMT p oposals in ol e much less ambi ious ep esen a ion schemas,
in pa icula ligh ly anno a ed ex in which wo ds a e accompanied by POS ags
and/o he esul s o “s emming” (i.e. mo phological analysis o iden i y oo o s em,
and pa ially in e p e endings). Planas (1999:8) illus a es he idea by conside ing
sen ence (1a) compa ed o each o (1b-d): al hough (1c) di e s by only one cha ac e ,
humans ins inc i ely ind (1b) a be e ma ch.
(1) a. The whi e ho se is nice.
b. The whi e ho ses a e nice.
c. The whi e house is nice.
d. The whi e houses a e nice.
Figu e 2. S uc u ed ep esen a ion sugges ed o ‘Click a colo and p ess ENTER.’
( om Planas and Fu use, 1999, p. 333).
An impo an ecen end in EBMT is o s o e simila examples in a uni ied
“gene alized” manne . Fo ins ance, he wo examples in (2a,b) could be gene alized
as (2c), and s o ed as such, wi h ob ious epe cussions o ma ching (see nex sec ion).
(2) a. John Mille lew o F ank u on Decembe 3 d.
b. D Howa d Johnson lew o I haca on 7 Ap il 1997.
c. <pe son-m> lew o <ci y> on <da e> .
An ea ly p oposal along hese lines is ound in Fu use and Iida’s (1992)
dis inc ion be ween “li e al examples” and “pa e n examples”, he la e con aining
a iables in place o wo ds, wi h he a iables cha ac e ized by a (lis o ) ypical
ille (s), as in (3).
8
8
They also ha e a hi d ype, called “g amma examples” which consis almos en i ely o a iables,
a he like he ules in con en ional MT.
(3) a. X o onegai shimasu → may I speak o he X′ (X=jimukyoku ‘o ice’, …)
b. X o onegai shimasu → please gi e me he X′ (X=bangō ‘numbe ’, …)
The idea is qui e widesp ead in he EBMT li e a u e, including Kaji e al.’s
(1992) “pseudo-sen ences”, Langé e al.’s (1997) “skele on sen ences” and a numbe
o o he s.
9
The examples a e usually gene alized by me ging simila cases, hough
au ho s di e as o whe he his can be done (semi-)au oma ically, o manually.
Ce ainly, i he idea was applied o TMSs, TMs could be much educed in size,
hough access would p esumably ha e o be mo e sophis ica ed.
3.3 Ma ching echniques
Accessing he TM o example-base in ol es “ma ching” he gi en ansla ion uni
agains he cases al eady s o ed. Ea ly implemen a ions o TMSs could handle only
exac ma ches, al hough alphanume ic “ eplaceables” as in (4) we e allowed.
(4) a. This is shown as A in he diag am.
b. This is shown as B in he diag am
Bowke in oduces he dis inc ion be ween an “exac ” ma ch and a “ ull” ma ch.
An exac ma ch is 100 pe cen iden ical o he segmen ha he ansla o is cu en ly ansla ing,
bo h linguis ically and in e ms o o ma ing. ... This means ha he wo s ings mus be
iden ical in e e y way, including spelling, punc ua ion, in lec ion, numbe s, and e en o ma ing
(e.g., i alics, bold). (Bowke , 2002:96 )
Thus (5b) would no be e ie ed as an exac ma ch o (5a) because o he di e ence
in o ma ing.
(5) a. Click on OK
b. Click on OK.
A “ ull ma ch” on he o he hand,
… occu s when a new sou ce segmen di e s om a s o ed TM uni only in e ms o so-called
a iable elemen s, which a e some imes e e ed o as “placeables” o “named en i ies”.
Va iable elemen s include numbe s, da es, imes, cu encies, measu emen s, and some imes
p ope names. (ibid., p. 98)
Bowke ’s “placeables” ha e been e med “ answo ds” by Gaussie e al.
(1992), while Macklo i ch and Russell (2000) call hem “non- ansla ables”. The
no ion o “named en i ies” is ound in In o ma ion Re ie al. As he la e au ho s
poin ou , hey a e ea ed in ansla ion in a a he anspa en manne , ei he no
ansla ed a all, o subjec o speci ic con en ions, and in any case, independen o
con ex . Fo a TMS hei impac is wo old. On a simple le el, we wan o ha e
ma che s ha “igno e” hem, so ha (4a,b) abo e a e e ec i ely “exac ma ches”.
Addi ionally, in a mo e sophis ica ed TMS (and in EBMT), hey a e impo an
building blocks o sugges ing au oma ically a likely ansla ion o he gi en ex .
Conside (6a) as a ex o be ansla ed, which ma ches wi h (6b), wi h he di e ences
highligh ed, and i s associa ed ansla ion (6c). In (6b) he e a e wo answo ds he
“ ansla ion” o which can be eadily iden i ied in (6c), bu you ha e o know he
a ge language o know which wo d o change in (6c) o accommoda e he lexical
di e ence la ge s. small.
9
Nomiyama (1992), Almuallim e al. (1994), Akiba e al. (1995), Collins and Cunningham (1995), Jain
e al. (1995), Ma sumo o and Ki amu a (1995), Wa anabe and Takeda (1998), Ca l (1999) – see Some s
(1999:139 ).
elemen s ma ch up: his can be done by looking a u he examples which isola e he
wo ds in ques ion.
O cou se much o his wo k could be simpli ied wi h he help o an on-line
dic iona y, assuming we had one. Bu wha is o in e es o esea che s is he ex en o
which i can be au oma ed. To exempli y his, conside he examples in (19), in a
language p obably un amilia o mos eade s.
(19) a. Dia nak pě gi kě kědai běli o i.
She is going o go o he shops o buy b ead.
b. Dia pě gi kě pasa nak běli baju.
She wen o he ma ke o buy a shi .
c. Mě eka pě gi kě kampung nak běli ke e a.
They wen o he illage o buy a ca .
I is no di icul o iden i y he p obable wo d-pai ings and om he examples o
cons uc he co ec ansla ions o sen ences like hose in (20), and we in i e he
eade o y i as an exe cise.
15
In doing so i should be no ed how much gene ic
(common-sense) knowledge abou how languages wo k we as humans b ing o his
ask, which may ha e o be simula ed in an o he wise pu ely au oma ic sys em.
(20) a. She wen o he illage o buy b ead.
b. They a e going o he ma ke .
5 TMSs would be be e i hey we e mo e like EBMT
Ou pu pose in his pape has been o poin ou how some o he ideas de eloped in
connec ion wi h EBMT could be in oduced in o he de elopmen o TMSs. In his
inal sec ion we a emp o summa ize he main p oposals.
• I hey could iden i y wha in he a ge pa o he ma ch has o be changed.
I we wan o be able o ma ch simila sen ences, such as hose di e ing only in he o m o hei
wo ds, we need linguis ic analysis. We do no need a deep analysis ha would ake a long ime
o p ocess, bu jus a “s emming” and a “ agging” one ha would gi e a ligh bu c ucial
analysis (Planas 1999:9).
[T]he mos p omising s a egy o he nex gene a ion o TM sys ems will be o employ a ious
pa ial pa sing o “chunking” echniques. (Macklo i ch, 2000)
• I hey could make sugges ions abou wha in he a ge pa o he ma ch has o
be changed o.
• I hey could cons uc a ge ex s om ma ched agmen s.
• I hey could ake simila examples and make gene aliza ions abou hem.
TM and EBMT can be seen o lie a opposi e ends o a spec um in memo y-based ansla ion.
On he one hand, TM equi es ew linguis ic[ ] esou ces bu canno combine agmen s om
di e en T[ ansla ion] U[ni ]s, and on he o he hand, EBMT can combine example agmen s,
bu does so by elying on … knowledge-in ensi e ools. (McTai e al. 1999)
15
The language is Malay. The co ec ansla ions a e Dia pě gi kě kampung nak běli o i; Mě eka nak
pě gi kě pasa .

Re e ences
Ah enbe g, L. and M. Me kel: 1996. ‘On T ansla ion Co po a and T ansla ion Suppo Tools: A
P ojec Repo ’, in K. Aijme , B. Al enbe g, and M. Johansson (eds), Languages in Con as ,
Pape s om a Symposium on Tex -based C oss-linguis ic S udies, Lund: Lund Uni e si y P ess,
pp. 185–200.
Akiba, Y., M. Ishii, H. Almuallim and S. Kaneda: 1995. ‘Lea ning English Ve b Selec ion Rules om
Hand-made Rules and T ansla ion Examples’, in P oceedings o he Six h In e na ional
Con e ence on Theo e ical and Me hodological Issues in Machine T ansla ion, Leu en,
Belgium, pp. 206–220.
Al-Adhaileh, M. H. and Tang E. K.: 1998. ‘A Flexible Example-Based Pa se Based on he SSTC’, in
COLING-ACL ‘98: 36 h Annual Mee ing o he Associa ion o Compu a ional Linguis ics and
17 h In e na ional Con e ence on Compu a ional Linguis ics, Mon eal, Quebec, pp. 687–693.
Almuallim, H., Y. Akiba, T. Yamazaki, A. Yokoo and S. Kaneda: 1994. ‘Two Me hods o Lea ning
ALT-J/E T ansla ion Rules om Examples and a Seman ic Hie a chy’, in COLING 94, The 15 h
In e na ional Con e ence on Compu a ional Linguis ics, Kyo o, Japan, pp. 57–63.
Aus e mühl, F. 2001. Elec onic Tools o T ansla o s. Manches e : S . Je ome.
Anon: 1991. ‘Bi ex makes p og ess wi h and wi hou he name’, Language In e na ional 3.2:5–6.
A he n, P. J.: 1978. ‘Machine T ansla ion and Compu e ized Te minology Sys ems: A T ansla o ’s
Viewpoin ’, in B.M. Snell (ed.) T ansla ing and he Compu e : P oceedings o a Semina ,
London, 14 h No embe 1978, Ams e dam (1979): No h Holland, pp. 77–108.
A he n, P. J.: 1981. ‘Aids Unlimi ed: The Scope o Machine Aids in a La ge O ganiza ion’, Aslib
P oceedings, 33:309–319.
Benne , P. 1994. ‘The T ansla ion Uni in Human and Machine’, Babel 40:12–20.
Bowke , L. 2002. Compu e -Aided T ansla ion Technology. A P ac ical In oduc ion. O awa:
Uni e si y o O awa P ess.
B ace, C.: 1992. ‘The Many Fla o s o T ansla ion Memo y’, Language Indus y Moni o (May/June
1992); a ailable a h p://www.lim.nl/moni o /memo y.h ml.
Ca l, M.: 1999. ‘Inducing T ansla ion Templa es o Example-Based Machine T ansla ion’, in Machine
T ansla ion Summi VII, Singapo e, pp. 250–258.
Ca l, M.: 2001. ‘Inducing T ansla ion G amma s om B acke ed Alignmen s’, in MT Summi VIII
Wo kshop on Example-Based Machine T ansla ion, San iago de Compos ela, Spain, pp. 12–23;
ep . in Ca l and Way (2003), pp. 339–361.
Ca l, M. and A. Way (eds): 2003. Example-Based Machine T ansla ion, Do d ech : Kluwe .
Ca e, J.R.: 1988. ‘Obse a ions on Bi-Tex ’, Le e o he Edi o , Language Mon hly 57:18–19.
Cha e jee, N.: 2001. ‘A S a is ical App oach o Simila i y Measu emen Be ween Sen ences o
EBMT’, in STRANS–2001 Symposium on T abnsla ion Suppo Sys ems, Kanpu , India. [no
page numbe s]
Cicekli, I., and H. A. Gü eni : 1996. ‘Lea ning T ansla ion Rules F om A Bilingual Co pus’, in
NeMLaP-2: P oceedings o he Second In e na ional Con e ence on New Me hods in Language
P ocessing, Anka a, Tu key, pp. 90–97; ep . as ‘Lea ning T ansla ion Templa es om
Bilingual T ansla ion Examples’ in Ca l and Way (2003), pp. 255–286.
Collins, B.: 1998. Example-Based Machine T ansla ion: An Adap a ion-Guided Re ie al App oach.
PhD hesis, T ini y College, Dublin.
Collins, B. and P. Cunningham: 1995. ‘A Me hodology o Example Based Machine T ansla ion’, in
CSNLP 1995: 4 h Con e ence on he Cogni i e Science o Na u al Language P ocessing,
Dublin.
Collins, B. and P. Cunningham: 1996. ‘Adap a ion-Guided Re ie al in EBMT: A Case-Based
App oach o Machine T ansla ion’, in I. Smi h and B. Fal ings (eds), Ad ances in Case-Based
Reasoning: Thi d Eu opean Wo kshop, EWCBR-96, Be lin: Sp inge , pp. 91–104.
C anias, L., H. Papageo giou and S. Pipe idis: 1997. ‘Example Re ie al om a T ansla ion Memo y’,
Na u al Language Enginee ing 3:255–277.
Denne , G.: 1995. T ansla ion memo y: Concep , p oduc s, impac and p ospec s. MSc disse a ion,
School o Elec ical, Elec onic and In o ma ion Enginee ing, Sou h Bank Uni e si y, London.
Esselink, B. 2000. A P ac ical Guide o Localiza ion, 2nd ed. Ams e dam: John Benjamins.
F ede king, R., S. Ni enbu g, D. Fa well, S. Helm eich, E. Ho y, K. Knigh , S. Beale, C. Domashne ,
D. A a do, D. G annes and R. B own: 1994. ‘In eg a ing T ansla ions om Mul iple Sou ces
wi hin he Pangloss Ma k III Machine T ansla ion Sys em’, in Technology Pa ne ships o
C ossing he Language Ba ie : P oceedings o he Fi s Con e ence o he Associa ion o
Machine T ansla ion in he Ame icas, Columbia, Ma yland, pp. 73–80.
F eibo , G.P.: 1992. ‘Compu e Aided T ansla ion in an In eg a ed Documen P oduc ion P ocess:
Tools and Applica ions’, in T ansla ing and he Compu e 14: Quali y S anda ds and he
Implemen a ion o Technology in T ansla ion, London, pp. 15–24.
Fu use, O. and H. Iida: 1994. ‘Cons i uen Bounda y Pa sing o Example-Based Machine T ansla ion’,
in COLING 94, The 15 h In e na ional Con e ence on Compu a ional Linguis ics, Kyo o, Japan,
pp. 105–111.
Gaussie , E., J.-M. Langé and F. Meunie : 1992. ‘Towa ds bilingual e minology’, in P oceedings o
he ALLC/ACH Con e ence, Ox o d, pp. 121–124.
Ge lo , P.: 1987. ‘Iden i ying he Uni o Analysis in T ansla ion’, in C. Fæ ch and G. Kaspe (eds),
In ospec ion in Second Language Resea ch, Cle edon: Mul ilingual Ma e s, pp. 135–158.
Gü eni , H. A. and I. Cicekli: 1998. ‘Lea ning T ansla ion Templa es om Examples’, In o ma ion
Sys ems 23:353–363.
Ha is, B.: 1988. ‘Bi- ex , a new concep in ansla ion heo y’, Language Mon hly 54:8–10.
Heyn, M. 1998: ‘T ansla ion Memo ies – Insigh s & P ospec s’, in L. Bowke , M. C onin, D. Kenny
and J. Pea son (eds) Uni y in Di e si y? Cu en T ends in T ansla ion S udies, Manches e : S
Je ome, pp. 123–136.
Hu chins, J. 1998: ‘The O igins o he T ansla o ’s Wo ks a ion’. Machine T ansla ion, 13:287–307.
Jain, R., R. M. K. Sinha and A. Jain: 1995. ‘Role o Examples in T ansla ion’, in 1995 IEEE
In e na ional Con e ence on Sys ems, Man and Cybe ne ics, Vancou e , BC, pp. 1615–1620.
Kaji, H., Y. Kida and Y. Mo imo o: 1992. ‘Lea ning T ansla ion Templa es om Bilingual Tex ’, in
P oceedings o he i een h [sic] In e na ional Con e ence on Compu a ional Linguis ics, Ac es
du quinzième colloque in e na ional en linguis ique in o ma ique, COLING-92, Nan es, F ance,
pp. 672–678.
Kay, M.: 1980. The P ope Place o Men and Machines in Language T ansla ion. Resea ch Repo
CSL-80-11, Xe ox PARC, Palo Al o, Cali . Rep in ed in Machine T ansla ion 12:3–23 (1997).
Keck, B.: 1989. ‘Theo e ical S udy o a S a is ical App oach o T ansla ion’, TWB Technical Repo ,
F aunho e Ins i u e IAO, S u ga , Ge many.
Ki ano, H.: 1993. ‘A Comp ehensi e and P ac ical Model o Memo y-Based Machine T ansla ion’, in
P oceedings o he In e na ional Join Con e ence on A i icial In elligence, Chambé y, F ance,
pp. 1276–1282.
Kugle , M., G. Heye , R. Kese, B. on Kleis -Re zow and G. Winkelmann: 1991. ‘The T ansla o ’s
Wo kbench: An En i onmen o Mul i-Lingual Tex P ocessing and T ansla ion’, in Machine
T ansla ion Summi III P oceedings, Washing on, DC, pp. 81–83.
Langé, J.-M., É. Gaussie and B. Daille: 1997. ‘B icks and Skele ons: Some Ideas o he Nea Fu u e
o MAHT’, Machine T ansla ion 12:39–51.
Le-Hong, K., M. Höge and A. Hohman: 1992. ‘Use ’s Poin o View o he T ansla o ’s Wo kbench’,
in T ansla ing and he Compu e 14: Quali y S anda ds and he Implemen a ion o Technology
in T ansla ion, London, pp. 25–32.
Macdonald, K.: 2001. ‘Imp o ing Au oma ic Alignmen o T ansla ion Memo y C ea ion’, in
T ansla ing and he Compu e 23: P oceedings om he Aslib Con e ence held on 29 & 30
No embe 2001, London, [pages no numbe ed].
Macklo i ch, E.: 2000. ‘Two Types o T ansla ion Memo y’, in T ansla ing and he Compu e 22:
P oceedings om he Aslib Con e ence 16 & 17 No embe 2000, London, [pages no
numbe ed].
Macklo i ch, E. and G. Russell: 2000. ‘Wha ’s Been Fo go en in T ansla ion Memo y’, in J. S. Whi e
(ed.) En isioning Machine T ansla ion in he In o ma ion Fu u e: 4 h
Con e ence o he
Associa ion o Machine T ansla ion in he Ame icas, AMTA 2000, Be lin: Sp inge , pp. 137–
146.
Manning, C. D. and H. Schü ze: 1999. Founda ions o S a is ical Na u al Language P ocessing.
Camb idge, Massachuse s: The MIT P ess.
Ma uyama, H. and H. Wa anabe: 1992. ‘T ee Co e Sea ch Algo i hm o Example-Based
T ansla ion’, in Fou h In e na ional Con e ence on Theo e ical and Me hodological Issues in
Machine T ansla ion. Empi icis s. Ra ionalis Me hods in MT. TMI-92, Mon éal, Québec, pp.
173–184.
Ma sumo o, Y., H. Ishimo o and T. U su o: 1993. ‘S uc u al Ma ching o Pa allel Tex s’, in 31s
Annual Mee ing o he Associa ion o Compu a ional Linguis ics, Columbus, Ohio, pp. 23–30.
Ma sumo o, Y. and M. Ki amu a: 1995. ‘Acquisi ion o T ansla ion Rules om Pa allel Co po a’, in R.
Mi ko and N. Nicolo (eds) Recen Ad ances in Na u al Language P ocessing: Selec ed
Pape s om RANLP’95, Ams e dam: John Benjamins, pp. 405–416.
McTai , K.: 2001. ‘Linguis ic Knowledge and Complexi y in an EBMT Sys em Based on T ansla ion
Pa e ns’, in MT Summi VIII Wo kshop on Example-Based Machine T ansla ion, San iago de
Compos ela, Spain, pp. 23–34; o be ep . Ca l and Way (2003).
McTai , K., M. Olohan and A. T ujillo: 1999. ‘A Building Blocks App oach o T ansla ion Memo y’,
in T ansla ing and he Compu e 21: P oceedings om he Aslib Con e ence, London, [pages
no numbe ed].
Melby, A. K.: 1995. The Possibili y o Language: A Discussion o he Na u e o Language.
Ams e dam: John Benjamins.
Melby, A. K.: 1998. ‘Behind he Scenes: Da a-exchange s anda ds a e unsung he oes e olu ionizing
he language indus ies’, Language In e na ional 10.6:30–31, 43.
Melby, A. K.: 2000. ‘Sha ing o ansla ion memo y da abases de i ed om aligned pa allel ex ’, in
Vé onis (2000), pp. 347–368.
Me kel, M., B. Nilsson and L. Ah enbe g: 1994. ‘A ph ase- e ie al sys em based on ecu ence’, in
Second Annual Wo kshop on Ve y La ge Co po a (WVLC2), Kyo o, Japan, pp. 99–108.
Nagao, M.: 1984. ‘A F amewo k o a Mechanical T ansla ion be ween Japanese and English by
Analogy P inciple’, in A. Eli ho n and R. Bane ji (eds) A i icial and Human In elligence,
Ams e dam: No h-Holland, pp. 173–180.
Ni enbu g, S., C. Domashne and D. J. G annes: 1993. ‘Two App oaches o Ma ching in Example-
Based Machine T ansla ion’, in P oceedings o he Fi h In e na ional Con e ence on
Theo e ical and Me hodological Issues in Machine T ansla ion TMI ‘93: MT in he Nex
Gene a ion, Kyo o, Japan, pp. 47–57.
Nomiyama, H.: 1992. ‘Machine T ansla ion by Case Gene aliza ion’, in P oceedings o he i een h
[sic] In e na ional Con e ence on Compu a ional Linguis ics, COLING-92, Nan es, F ance, pp.
714–720.
O’B ien, S. 1998: ‘P ac ical Expe ience o Compu e -Aided T ansla ion Tools in he So wa e
Localiza ion Indus y’, in L. Bowke , M. C onin, D. Kenny and J. Pea son (eds) Uni y in
Di e si y? Cu en T ends in T ansla ion S udies, Manches e : S Je ome, pp. 115–122.
Pappegaaij, B. C., V. Sadle and A. P. M. Wi kam (eds): 1986a. Wo d Expe Seman ics: An
In e lingual Knowledge-Based App oach. Do d ech : Reidel.
Pappegaaij, B. C., V. Sadle and A. P. M. Wi kam: 1986b. ‘Expe imen s wi h an MT-Di ec ed Lexical
Knowledge Bank’, in 11 h In e na ional Con e ence on Compu a ional Linguis ics:
P oceedings o Coling ’86, Bonn, Wes Ge many, pp. 432–434.
Planas, E.: 1999. ‘Towa ds Second-Gene a ion T ansla ion Memo ies’, LISA Newsle e 8.2:7–11.
Planas, E. and O. Fu use: 1999. ‘Fo malizing T ansla ion Memo ies’, in Machine T ansla ion Summi
VII, Singapo e, pp. 331–339; ep . in Ca l and Way (2003), pp. 157–188.
Rapp, R. 2002. ‘A Pa -o -Speech-Based Sea ch Algo i hm o T ansla ion Memo ies’ in LREC 2002,
Thi d In e na ional Con e ence on Language Resou ces and E alua ion, Las Palmas de G an
Cana ia, Spain, pp. 466–472.
Riesbeck, C. and R. Schank: 1989. Inside Case-Based Reasoning. Hillsdale, NJ: Law ence E lbaum.
Sadle , V.: 1991. ‘The Tex ual Knowledge Bank: Design, Cons uc ion, Applica ions’, in In e na ional
Wo kshop on Fundamen al Resea ch o he Fu u e Gene a ion o Na u al Language
P ocessing (FGNLP), Kyo o, Japan, pp. 17–32.
Sadle , V. and R. Vendelmans: 1990. ‘Pilo Implemen a ion o a Bilingual Knowledge Bank’, in
COLING-90, Pape s P esen ed o he 13 h In e na ional Con e ence on Compu a ional
Linguis ics, Helsinki, Finland, Vol. 3, pp. 449–451.
Sa o, S. and M. Nagao: 1990. ‘Towa d Memo y-Based T ansla ion’, in COLING-90, Pape s P esen ed
o he 13 h In e na ional Con e ence on Compu a ional Linguis ics, Helsinki, Finland, Vol. 3,
pp. 247–252.
Schäle , R., A. Way and M. Ca l: 2003. ‘EBMT in a con olled en i onmen ’, in Ca l and Way (2003),
pp. 83–114.
Schube , K.: 1986. ‘Linguis ic and Ex a-Linguis ic Knowledge: A Ca alogue o Language- ela ed
Rules and hei Compu a ional Applica ion in Machine T ansla ion’, Compu e s and T ansla ion
1:125–152.
Sibley, J.: 1988. ‘Le sys ème ALPS’, in Ac es du Séminai e in e na ionale “La aduc ion assis ée pa
o dina eu ”, Pe spec i es echnologiques, indus ielles e économiques en isageables à
l’ho izon 1990”, Pa is, pp. 95–102.
Sima d, M.: 2003. Mémoi es de aduc ion sous-ph as iques. Thèse de doc o a , Uni e si é de
Mon éal.
Sima d, M. and P. Langlais: 2001. ‘Sub-sen en ial Exploi a ion o T ansla ion memo ies’, in MT
Summi VIII: Machine T ansla ion in he In o ma ion Age, San iago de Compos ela, Spain, pp.
335–339.
Some s, H.: 1999. ‘Re iew A icle: Example-based Machine T ansla ion’, Machine T ansla ion
14:113–157; e ised e sion in Ca l and Way (2003), pp. 3–57.
Some s, H., I. McLean and D. Jones: 1994. ‘Expe imen s in Mul ilingual Example-Based Gene a ion’,
in CSNLP 1994: 3 d Con e ence on he Cogni i e Science o Na u al Language P ocessing,
Dublin, I eland, [pages no numbe ed].
Some s, H., J. Tsujii and D. Jones: 1990. ‘Machine T ansla ion wi hou a Sou ce Tex ’, in COLING-90,
Pape s P esen ed o he 13 h In e na ional Con e ence on Compu a ional Linguis ics, Helsinki,
Finland, Vol. 3, pp. 271–276.
Sumi a, E. and H. Iida: 1995. ‘He e ogeneous Compu ing o Example-Based T ansla ion o Spoken
Language’, in P oceedings o he Six h In e na ional Con e ence on Theo e ical and
Me hodological Issues in Machine T ansla ion, Leu en, Belgium, pp. 273–286.
Sumi a, E., H. Iida and H. Kohyama: 1990. ‘T ansla ing wi h Examples: A New App oach o Machine
T ansla ion’, in The Thi d In e na ional Con e ence on Theo e ical and Me hodological Issues
in Machine T ansla ion o Na u al Language, Aus in, Texas, pp. 203–212.
Sumi a, E. and Y. Tsu sumi: 1988. ‘A T ansla ion Aid Sys em Using Flexible Tex Re ie al Based on
Syn ax-Ma ching’, in Second In e na ional Con e ence on Theo e ical and Me hodological
Issues in Machine T ansla ion o Na u al Languages, P oceedings Supplemen , Pi sbu gh,
Pennsyl ania, [pages no numbe ed].
S anholm, F.: 1992. ‘“The Happy T iad” – The Human, he MAT, and he MT’, in T ansla ing and he
Compu e 14: Quali y S anda ds and he Implemen a ion o Technology in T ansla ion, London,
pp. 15–24.
Topping, S.: 2000. ‘Sha ing T ansla ion Da abase In o ma ion: Conside a ions o de eloping an
e hical and iable exchange o da a’, Mul iLingual Compu ing & Technology 11.5:59–61.
Veale, T. and A. Way: 1997. ‘Gaijin: A Boo s apping App oach o Example-Based Machine
T ansla ion’, In e na ional Con e ence, Recen Ad ances in Na u al Language P ocessing,
Tzigo Cha k, Bulga ia, pp. 239–244.
Vé onis, J. (ed.): 2000. Pa allel ex p ocessing: Alignmen and Use o T ansla ion Co po a.
Do d ech : Kluwe .
Wa anabe, H.: 1994. ‘A Me hod o Dis inguishing Excep ional and Gene al Examples in Example-
Based T ans e Sys ems’, in COLING 94, The 15 h In e na ional Con e ence on Compu a ional
Linguis ics, Kyo o, Japan, pp. 292–301.
Wa anabe, H. and K. Takeda: 1998. ‘A Pa e n-Based Machine T ansla ion Sys em Ex ended by
Example-based P ocessing’, in COLING-ACL ’98: 36 h Annual Mee ing o he Associa ion o
Compu a ional Linguis ics and 17 h In e na ional Con e ence on Compu a ional Linguis ics,
Mon eal, Quebec, pp. 1369–1373.
Way, A.: 2001. ‘T ansla ing wi h Examples’, in MT Summi VIII Wo kshop on Example-Based
Machine T ansla ion, San iago de Compos ela, Spain, pp. 66–80; ep . in Ca l and Way (2003),
pp. 443–472.
Wea e , A.: 1988. ‘Two Aspec s o In e ac i e Machine T ansla ion’, in M. Vasconcellos (ed.)
Technology as T ansla ion S a egy, Bingham on, NY: S a e Uni e si y o New Yo k a
Bingham on (SUNY), pp. 116–123.
Wu, D.: 2000a. ‘Alignmen ’, in R. Dale, H. Moisl and H. Some s (eds) Handbook o Na u al Language
P ocessing, New Yo k: Ma cel Dekke Inc., pp. 415–458.
Wu, D.: 2000b. ‘B acke ing and aligning wo ds and cons i uen s in pa allel ex using S ochas ic
In e sion T ansduc ion G amma s’, in Vé onis (2000), pp. 139–167.
Zhao, G. and J. Tsujii: 1999. ‘T ans e in Expe ience-Guided Machine T ansla ion’, in Machine
T ansla ion Summi VII, Singapo e, pp. 501–508.