Chap e 5
Teaching ansla ion wi h AI: B idging
heo y and p ac ice h ough p omp
enginee ing
Masa u Yamada
Rikkyo Uni e si y, Japan
This chap e explo es he inno a i e applica ion o la ge language models (LLMs)
in ansla o aining, ocusing on he use o ew-sho p omp s and chain-o -
hough p omp ing. I p oposes a no el app oach ha in eg a es me alanguages
and concep s o T ansla ion S udies in o p omp enginee ing, mo ing beyond
adi ional na u al language p ocessing goals o imp o ing machine ansla ion
quali y. The chap e demons a es how his me hod can c ea e in e ac i e and
engaging lea ning expe iences o ansla ion s uden s, allowing hem o explo e
a ious ansla ion s a egies and de elop c i ical hinking skills. Th ough con-
c e e examples, he chap e illus a es he po en ial o LLMs o gene a e di e se
ansla ion a ia ions and p o ide insigh ul analyses o ansla ion p ocesses.
While acknowledging limi a ions and he need o c i ical e alua ion, he esea ch
emphasises he posi i e and p oac i e possibili ies o LLMs in ansla o aining.
This app oach no only b idges he gap be ween ansla ion heo y and p ac ice
bu also opens new a enues o au onomous lea ning and he de elopmen o
essen ial skills o u u e ansla o s in he AI e a.
1 In oduc ion
Recen ad ancemen s in a i icial in elligence (AI), pa icula ly he eme gence o
La ge Language Models (LLMs), ha e gene a ed ex ensi e deba e ega ding hei
bene i s and d awbacks. In esponse, he UK’s Ins i u e o T ansla ion and In e -
p e ing (ITI) has a icula ed he Slow T ansla ion Mani es o (ITI 2024). D awing
Masa u Yamada. 2026. Teaching ansla ion wi h AI: B idging heo y and p ac ice
h ough p omp enginee ing. In JC Pene , Joss Moo kens & Masa u Yamada (eds.),
Teaching ansla ion in he age o gene a i e AI: New pa adigm, new lea ning?, 87–104.
Be lin: Language Science P ess. DOI: 10.5281/zenodo.17641072
Masa u Yamada
on an analogy be ween as ood and AI-gene a ed ansla ions, his mani es o
highligh s he po en ial socie al and p o essional ha ms o p io i ising speed o e
quali y in ansla ion p ac ices. I ad oca es o a enewed app ecia ion o slow,
human-led ansla ion, emphasising i s a is y, e hical igou , and c oss-cul u al
compe ence – quali ies ha a e o en sac i iced in apid, machine-d i en ans-
la ion p ocesses.
The Socié é ançaise des aduc eu s (SFT), he F ench ansla o s’ associa ion,
has also aised conce ns abou he inadequa e quali y o AI-based ansla ions
and he anspa ency o he esou ces used o machine lea ning in LLMs. The
SFT wa ns ha un e e ed access o AI ools, such as Cha GPT, could lead o
a decline in espec o language p o essionals o esul in an “unchecked ush
o ansla ions ha we a e asked o enhance” (Sla o 2024). Simila s a emen s
ega ding he po en ial c isis posed by AI echnology in ansla ion ha e been
issued by o he o ganisa ions, including CEATL (Eu opean Council o Li e a y
T ansla o s’ Associa ions 2024), ATA (Ame ican T ansla o s Associa ion 2024),
and JAT (Japan Associa ion o T ansla o s 2024).
The EU’s Language in he Human-Machine E a (LITHME) p ojec , ocusing on
language in he echnology e a, aims o “p epa e language esea che s o wha
is coming” and “ acili a e longe - e m dialogue be ween linguis s and echnology
de elope s” (LITHME P ojec 2021; see also Saye s e al. 2021). This comp ehen-
si e p ojec compiles expe opinions and iews, aiming o coun e ac he po en-
ial o e eliance on AI-based echnology and he ma ke -d i en o e selling o AI
(Saye s e al. 2021, Way 2025). By p omo ing such pe spec i es, he p ojec plays
a c ucial ole in encou aging a mo e cau ious and hough ul app oach owa ds
AI ad ancemen s. These in es iga ions and s a emen s om his p ojec se e
as a eminde o pause and c i ically e lec on he apidly accele a ing ends in
ou socie y.
Howe e , upon close examina ion o hese s a emen s and su eys, i becomes
appa en ha he unde s anding and p edic ions abou echnology a e no al-
ways accu a e o up- o-da e, and may no be en i ely e idence-based. Fo in-
s ance, while LITHME’s 2021 epo (Saye s e al. 2021) does ouch on concep s
ele an o LLMs and Gene a i e AI (GenAI), hese e e ences a e minimal and do
no encompass he ans o ma i e de elopmen s ha ha e since occu ed, such
as cha -based LLMs. This highligh s he challenge o keeping pace wi h he apid
e olu ion o echnology, which e en a ha ime was di icul o ully an icipa e.
Simila ly, he SFT s a es ha he ou pu o machine ansla ion (MT) emains
un eadable in i s aw s a e and equi es human co ec ion h ough pos -edi ing.
Howe e , i also adds ha “70 pe cen o ou membe ansla o s who esponded
o ou su ey conside ed PE (and by ex ension AI) a h ea o hei p o ession”
88
5 Teaching ansla ion wi h AI
(Sla o 2024). The con lic ing s a emen lea es unclea which e a o ype o MT i
e e s o, and he appa en emo ional eac ion e lec ing ea s o job loss sugges s
a lack o cla i y abou he in ended scope and echnological con ex o he claim.
This also implies ha such s a emen s may no ully e lec he cu en s a e o
cu ing-edge echnology.
In he con ex o ansla o aining, which is he p ima y ocus o his book,
he e may no be as s ong a backlash agains AI as in pa s o he indus y. Fo
ins ance, while some indus y s akeholde s see AI as a ool o enhance e iciency
and p o ide oppo uni ies o pos -edi ing wo k, many independen p o essional
ansla o s exp ess signi ican conce ns. Fea s abou he indisc imina e use o AI,
especially by un ained use s, pe sis due o i s po en ial o cause quali y issues
and unde mine he alue and compensa ion o human ansla o s. Acco ding
o ELIS (2025), a ound 64% o s uden s in ansla ion p og ammes epo using
GenAI a leas occasionally, wi h 19% using i egula ly. Uni e si y s a es ima e
MT use a 63%, compa ed o 58% among s uden s. These igu es ma k a s eep ise
om he limi ed adop ion epo ed in 2024 and highligh he con as be ween
he compa a i ely cau ious p o essional indus y and he apid in eg a ion o
GenAI in ansla o aining.
This chap e p oposes le e aging he accumula ed knowledge asse s o p o-
essional ansla o s and ansla ion esea che s o ac i ely in e ac wi h LLMs
h ough basic p omp s such as Chain o Though (CoT) p omp ing and ew-sho
p omp s. Fo example, ansla ion memo ies, which a e eposi o ies o pas hu-
man ansla ions pai ed wi h hei sou ce ex s, can be e ec i ely used o ans-
la o aining in conjunc ion wi h ew-sho p omp s. Addi ionally, he concep s
o ansla ion b ie s and wo k ins uc ions, adi ionally used in p o essional
ansla ion, can be di ec ly applied as p omp s. Fu he mo e, he classical asse s
o ansla ion esea ch ha desc ibe ansla ion s a egies can be adap ed o use
as CoT p omp ing.
The concep s s udied in ansla ion esea ch, which explain he ac o ansla-
ion, can be collec i ely e e ed o as he me alanguages o ansla ion (Miya a
e al. 2023). These me alanguages can be e ec i ely applied o p omp c ea ion.
These a emp s a e no me ely an enginee ing pe spec i e o enhance he accu-
acy o ansla ion ou pu s om LLMs bu a e p oposed as signi ican conside -
a ions o ansla o aining in he AI e a.
2 Li e a u e e iew
P e ious s udies on p omp s o ansla ion and LLMs ha e p ima ily been pub-
lished in he ield o na u al language p ocessing (NLP). Mos esea ch has o-
89
Masa u Yamada
cused on using LLMs as eplacemen s o exis ing MT sys ems o as ools o ad-
d ess he sho comings o MT. Speci ically, hese s udies o en in es iga e how
ansla ion accu acy compa es be ween LLM-gene a ed ansla ions and adi-
ional MT when using simple p omp s, such as ze o-sho p omp s ha me ely
ins uc he model o ansla e om one language o ano he . Fo example, se -
e al s udies (such as Hendy e al. 2023, Jiao e al. 2023, Wang e al. 2023) ha e
examined how accu a ely LLMs can ansla e using ze o-sho p omp s and ound
ha , o high- esou ce languages—languages wi h abundan digi al esou ces
and aining da a such as English, Spanish, o Chinese—LLMs can p oduce ans-
la ions wi h accu acy compa able o o exceeding ha o adi ional MT (Way
2025).
O he s udies ha e explo ed using LLMs and p omp s o enhance ansla ion
quali y, pa icula ly o low- esou ce languages. Fo ins ance, Jiao e al. (2023)
analysed Cha GPT’s MT capabili ies and ound ha while i compe es wi h com-
me cial sys ems o high- esou ce Eu opean languages, i s uggles wi h low-
esou ce and dis an languages. Gao e al. (2023) de eloped a new me hod o
ansla ion p omp s ha includes ask in o ma ion, con ex ual domain in o ma-
ion, and pa -o -speech ags, which signi ican ly imp o ed Cha GPT’s pe o -
mance, su passing comme cial sys ems in mul iple ansla ion di ec ions. Zhang
e al. (2023) p o ided a comp ehensi e summa y o p omp s a egies used o
da e and a emp ed o o e come he sho comings in ansla ion o low- esou ce
language combina ions and o he challenging scena ios.
Recen expe imen s ha e also ocused on he ansla ion o high-con ex and
mul i-modal ma e ials such as Japanese manga using LLMs. Yang e al. (2024)
conduc ed expe imen s demons a ing ha applying app op ia e p omp s can
signi ican ly imp o e ansla ion accu acy in such con ex s. Howe e , p ac ical
ansla ion o ganisa ions ha e aised conce ns abou he easibili y o applying
hese me hods o manga, pa icula ly in cap u ing i s nuanced cul u al and con-
ex ual laye s, which a e in eg al o he gen e (Japan Associa ion o T ansla o s
2024).
While hese s udies o igina e om he NLP ield and end o ocus on engi-
nee ing imp o emen s, he e appea s o be a di e ence in he unde s anding o
ansla ion be ween NLP esea che s and T ansla ion S udies (TS) schola s. In
NLP, ansla ion is o en app oached – pa icula ly in e alua ion – as i he e
we e a single co ec ou come. A he sys em le el, howe e , MT models may
gene a e mul iple possible ansla ions and hen ou pu he one judged o be
he “mos likely”, which does no necessa ily align wi h wha a human expe
would conside he “bes ”. By con as , TS gene ally ecognises ha ansla ion
may a y acco ding o con ex and pu pose. This a iabili y makes i di icul
90
5 Teaching ansla ion wi h AI
o de ine wha cons i u es a “good” o “quali y” ansla ion. Fo ins ance, classi-
cal ansla ion heo ies such as Skopos heo y sugges ha ansla ion s a egies
(e.g. domes ica ion s. o eignisa ion (Venu i 1995), co e s. o e ansla ion
(House 1981)) migh depend on he pu pose o he ansla ion. O e ime, TS has
de eloped insigh s in o hese a ia ions.
The ollowing sec ions a emp o add ess he gap le by NLP esea che s by
inco po a ing he me alanguages o TS, namely he concep s o ansla ion, in o
LLM p omp s and examining he esul ing changes in ansla ion ou pu .
Yamada (2023), o ins ance, in es iga ed he impac o inco po a ing ansla-
ion pu poses and a ge audiences in o p omp s on Cha GPT’s ansla ion ou -
pu . This s udy ocused on he p e-p oduc ion phase o he ansla ion p ocess,
d awing on p e ious ansla ion esea ch, indus y p ac ices, and ISO s anda ds.
By including concep s and e ms om TS as p omp s, he esea ch explo ed he
po en ial o achie ing lexible ansla ions ha adi ional MT sys ems ha e
s uggled o p oduce. The s udy e alua ed changes in ansla ion ou pu using
subjec i e quali a i e assessmen s and cosine simila i y, inco po a ing concep s
such as dynamic equi alence (Nida & Tabe 1969/2003).
Fu he , He (2024) explo ed using ansla ion esea ch concep s o design
p omp s o LLMs o imp o e ansla ion quali y. The s udy discussed he e -
ec i eness o inco po a ing concep ual ools and pe sonas o ansla o s and
au ho s in o Cha GPT’s ansla ion ask p omp s. Al hough he small-scale ex-
pe imen s indica ed limi ed e ec i eness in imp o ing ansla ion quali y, he
pape highligh ed he need o u he esea ch on he impac o TS concep s on
LLMs ansla ion asks.
Building on he ideas om hese wo pape s, his chap e p esen s he s udy
inding e alua ing he e ec i eness and po en ial o inco po a ing TS concep s
as p omp s o ansla o aining, a he han assessing imp o emen s in LLM-
gene a ed ansla ion quali y. I u he discusses he implica ions o his ap-
p oach o ansla ion educa ion.
3 Aims and pedagogical po en ial o his chap e
As obse ed in he compa a i e analysis o p e ious esea ch, p omp s in NLP
o en end o p io i ise he enginee ing aspec o achie ing equi alence be ween
sou ce and a ge ex s, ypically measu ed by s anda dised ansla ion me ics
such as BLEU (Papineni e al. 2002) and COMET (Rei e al. 2020). This app oach
gene ally ocuses on how closely he ansla ion aligns wi h he sou ce ex .
In con as , TS o en emphasises he impo ance o how he ansla ion is e-
cei ed by he a ge audience, conside ing he con ex and si ua ional a iables.
91
Masa u Yamada
This philosophical challenge in e alua ing ansla ion quali y in ol es nume -
ous a iables and may no be solely based on sou ce- a ge ex equi alence. By
inco po a ing he concep s and me alanguages o TS, i may be possible o c ea e
p omp s o LLMs ha be e add ess hese complex conside a ions.
This chap e aims o se e as a s a ing poin o designing pedagogical ac i i-
ies ha le e age concep s and e minologies (me alanguages) om TS o c ea e
e ec i e and unique p omp s o LLMs. While no a emp ing o exhaus i ely
e alua e all me alanguages as LLM p omp s o educa ional use, he chap e
p esen s se e al speci ic examples o inspi e ansla ion educa o s and ins uc-
o s in hei p ac ice. By doing so, i highligh s he po en ial o in eg a ing TS
knowledge in o LLM-assis ed ansla ion educa ion.
To achie e his, he chap e i s p o ides an o e iew o he ounda ional
concep s o p omp enginee ing o LLMs, speci ically ew-sho p omp s and CoT
p omp ing. I hen explo es how hese concep s migh be in eg a ed wi h exis ing
TS knowledge. Th ough hese examples and discussions, he chap e aims o o e
p ac ical sugges ions and po en ial di ec ions o inco po a ing TS concep s in o
LLM-assis ed ansla ion educa ion, a he han p esen ing de ini i e solu ions.
4 Few-Sho p omp s and CoT p omp ing
The concep s o ew-sho p omp s and CoT p omp ing a e explained in he
P omp Enginee ing Guidelines.1Applying hese concep s in his chap e e-
qui es some modi ica ions and expanded in e p e a ions, wi h conc e e examples
p o ided in Sec ions 5 and 6.
Few-sho p omp ing is a p omp design me hod o LLMs ha includes a ew
examples o inpu -ou pu pai s along wi h ask ins uc ions. This app oach al-
lows he model o in e he ask’s in en and ou pu o ma om he p o ided
examples, po en ially leading o mo e accu a e and desi able ou pu s. Few-sho
p omp ing is pa icula ly e ec i e o complex asks o when ze o-sho p omp -
ing (ins uc ions wi hou examples) may be insu icien .
Conside a scena io simila o a ansla ion memo y whe e po ions o sou ce
and a ge ex s a e p o ided o he LLM. The model may lea n implici pa e ns
om hese examples and apply hese lea ned pa e ns o gene a e ansla ions.
This me hod pa allels ac ual p ac ices in he ansla ion indus y, whe e ans-
la ion se ice p o ide s o en p o ide ansla ion memo ies o human ansla-
o s. T ansla o s ypically lea n om hese pas ansla ions, obse ing how s yle
guide ules a e conc e ely applied. Simila ly, ansla ion coo dina o s o p ojec
1h ps://www.p omp ingguide.ai
92
5 Teaching ansla ion wi h AI
manage s p o ide ins uc ions o human ansla o s. This chap e aims o in es-
iga e how compa able ins uc ions, when used as ew-sho p omp s o LLMs,
migh a ec ansla ion ou pu .
CoT p omp ing in ol es e balising he hough p ocess as a p omp . Jus as
humans ollow a s ep-by-s ep easoning p ocess o sol e p oblems, CoT p omp -
ing encou ages LLMs o explici ly ollow a CoT. This leads o mo e accu a e and
in e p e able esponses. In he con ex o ansla ion, ew-sho p omp s p o ide
examples o sou ce and a ge ex s wi hou explaining he in e media e p ocess.
In con as , CoT p omp ing could desc ibe he ansla ion p ocess, e e encing
ansla ion p ocess esea ch li e a u e. Fo ins ance, a ansla o migh ead he
sou ce ex , p oduce a li e al ansla ion, conside he con ex , moni o he ans-
la ion, and hen e ise i o make sense o he a ge audience, cul u e, and ead-
e s (e.g., moni o model, Ti kkonen-Condi 2005). This s ep-by-s ep desc ip ion
o he ansla ion p ocess can be gi en as a CoT p omp ing.
Addi ionally, as sugges ed by Yamada (2023), CoT p omp ing can be in e -
p e ed as de ailed ansla ion b ie s. He (2024) demons a ed ha se ing ans-
la o pe sonas (p o iles) can also be conside ed a a ian o CoT p omp ing. In
p o essional ansla ion se ings, companies ypically selec sui able ansla o s
o speci ic asks and p o ide de ailed ansla ion b ie s. This chap e in e p e s
CoT p omp ing as analogous o he language used o desc ibing ansla ion p o-
cesses (Miya a e al. 2023) and in es iga es how such p omp s migh in luence
ansla ion ou pu , as well as hei po en ial educa ional implica ions.
5 Example o a ew-sho p omp
Fi s ly, we p esen an example o a ew-sho p omp . This example aims o ex-
amine how he ou pu changes when a po ion o a ansla ion memo y is inpu
in o he LLM. Howe e , om an educa ional pe spec i e, i is aluable o con-
side how lea ne s, when aced wi h hei own ansla ion asks, can s udy and
emula e a ious linguis ic aspec s om he gi en co pus and inco po a e hem
in o hei subsequen ansla ions.
In his case, we delibe a ely ocused on linguis ic one and manne . A highly
dis inc i e ansla ion co pus was p epa ed. As indica ed in he p omp below,
he o iginal English ex was con ex ualised as a monologue e lec ion o an el-
de ly pe son on hei li e. In he co esponding Japanese ansla ion, we used
exp essions cha ac e ised by sen ence endings ypical o he linguis ic s yle as-
socia ed wi h Japanese elde ly indi iduals, c ea ing an exagge a ed ansla ion
ha exempli ies hei manne o speaking. Al hough his example is qui e ex-
93
Masa u Yamada
eme and may lack ealism, i can be likened o he speci ic s yle and oice used
in a pa icula company’s use manuals in a eal-wo ld con ex .
The pu pose o his app oach is wo old: o obse e how he LLM adap s i s
ou pu based on he p o ided ansla ion memo y, and o encou age lea ne s o
c i ically analyse how hey can lea n om and apply speci ic linguis ic ea u es,
ones, and s yles in hei own ansla ion wo k. This exe cise, while using an
exagge a ed example, aims o heigh en awa eness o he impo ance o adap ing
language o speci ic con ex s and audiences in ansla ion p ac ice.
This p epa ed sample se was hen gi en as a ew-sho p omp o he LLM,
ins uc ing i o ansla e a new sou ce ex ollowing he co pus. The LLM used
was Claude 3.5 Sonne . To p o ide a compa ison, he new sou ce ex was also
ansla ed using he MT se ice DeepL. Subsequen ly, he LLM ansla ions and
he MT ansla ions we e compa ed agains he o iginal co pus. Addi ionally,
e alua ions we e conduc ed using he au oma ic e alua ion ool COMET2The
esul s o hese compa isons and e alua ions a e p esen ed below.
LLM P omp 5.1:
Below a e p o ided he English ex [sou ce ex ], in which an elde ly
indi idual li ing in con empo a y Ame ica e lec s on hei li e, and i s
Japanese ansla ion [ a ge ex ]. Based on his co pus, please ansla e
he [new sou ce ex ] in o Japanese.
[sou ce ex ] “When I was young, li e was so di e en . We didn’ ha e
all hese gadge s and echnology. We had o wo k ha d o e e y hing
we had. Bu , hose we e good imes oo. I emembe walking miles o
school, playing ou side un il dusk, and e e yone knew each o he in he
communi y.”
[ a ge ex ] 「若い頃はのう、生活はまったく違っとったんじゃ。今みたい
なガジェットやテクノロジーはなかったんじゃ。何でも手に入れるために一生
懸命働かなければならんかったんじゃ。でも、それもまた良い時代だったんじ
ゃよ。学校まで何マイルも歩いて、夕暮れまで外で遊んで、みんながコミュニ
ティの中でお互いを知っとったんじゃ。」
2We used WMT20-COMET o his e alua ion.
94
5 Teaching ansla ion wi h AI
[new sou ce ex ] “I miss hose simple days. We didn’ ha e much, bu
we we e happy. I spen my summe s helping my a he on he a m and
lea ning abou li e om him. Sundays we e always special, wi h amily
ga he ings and big dinne s. As I g ew olde , he wo ld changed apidly,
and some imes i ’s ha d o keep up. Bu I che ish hose memo ies, hey
keep me g ounded.”
Figu e 1: Claude s. DeepL.
Acco ding o he COMET sco es in Figu e 1, Claude’s sco es ange om 0.360
o 0.995, while DeepL’s sco es ange om 0.633 o 1.051. These sco es indica e
gene ally high-quali y ansla ions, as sco es close o 1.0 ypically e lec s ong
alignmen wi h he e e ence ex . On a e age, DeepL achie es highe sco es a
0.626 compa ed o Claude a 0.499.
Howe e , when e alua ed by a human ansla o , a s a k s ylis ic di e ence
becomes e iden be ween he wo. Claude T ansla ion skil ully cap u es he dis-
inc i e elde ly speech pa e ns ound in he e e ence ansla ions. Using a small
co pus, i c ea es he imp ession ha he same elde ly pe son is con inuing he
con e sa ion seamlessly. In con as , DeepL T ansla ion employs a comple ely
s anda d Japanese one, making i appea as hough a di e en pe son is speak-
ing, he eby dis up ing he low o he monologue. This disc epancy highligh s a
limi a ion o he COMET sco ing sys em: i ails o accoun o s ylis ic ea u es
o cul u al nuances such as hose p esen in Japanese “elde ly speech pa e ns”.
95
Masa u Yamada
p oach may o e oppo uni ies o os e au onomous lea ning and de elop es-
sen ial quali ies o independen ansla o s.
I is also impo an o acknowledge some limi a ions. No all esponses om
LLMs we e accu a e, and some p omp s we e less success ul han o he s. Fo
ins ance, while he LLM co ec ly explained why COMET could no p o ide a
ai e alua ion in he ew-sho p omp example, i ailed o gi e a easonable
answe when asked which ansla ion (Claude o DeepL) was close o he e -
e ence ansla ion. Such e o s and limi a ions o LLMs become mo e appa en
wi h inc eased use. Howe e , gi en ha ansla o aining inhe en ly equi es
main aining a c i ical pe spec i e, I belie e explo ing he possibili ies o using
LLMs is as impo an as conside ing he isks.
In conclusion, his chap e has demons a ed conc e e me hods o explo ing
he po en ial o LLMs in ansla ion educa ion. By le e aging he concep s and
me alanguages o TS in p omp enginee ing, we can c ea e mo e engaging, in-
e ac i e, and e ec i e lea ning expe iences o ansla ion s uden s. While chal-
lenges and limi a ions exis , he po en ial bene i s o in eg a ing LLMs in o ans-
la o aining a e signi ican and wa an u he in es iga ion and de elopmen .
Re e ences
Ame ican T ansla o s Associa ion. 2024. ATA s a emen on a i icial in elligence.
h ps://www.a ane .o g/ad ocacy- ou each/a a- s a emen - on- a i icial-
in elligence/.
Do s , Ale a G. 2024. Me apho in li e a y machine ansla ion: S yle, c ea i -
i y and li e a iness. In And ew Ro hwell, Andy Way & Roy Youdale (eds.),
Compu e -assis ed li e a y ansla ion, 173–186. New Yo k, USA: Rou ledge.
DOI: 10.4324/9781003357391-9.
ELIS. 2025. ELIS 2025 Eu opean language indus y su ey. Tech. ep. ELIS Re-
sea ch. 1–53. h p://elis- su ey.o g/wp- con en /uploads/2025/03/ELIS-
2025_Repo .pd .
Eu opean Council o Li e a y T ansla o s’ Associa ions. 2024. No one le behind,
no language le behind, no book le behind. h ps://www.cea l.eu/no-one-le -
behind-no-language-le -behind-no-book-le -behind.
Gao, Yuan, Ruili Wang & Feng Hou. 2023. How o design ansla ion p omp s o
Cha GPT: an empi ical s udy. h ps://a xi .o g/abs/2304.02182.
He, Sui. 2024. P omp ing Cha GPT o ansla ion: a compa a i e analysis o ans-
la ion b ie and pe sona p omp s. h ps://a xi .o g/abs/2403.00127.
102
5 Teaching ansla ion wi h AI
Hendy, Am , Mohamed Abdel ehim, Am Sha a , Vikas Raunak, Mohamed
Gab , Hi okazu Ma sushi a, Young Jin Kim, Mohamed A i y & Hany Hassan
Awadalla. 2023. How good a e GPT models a machine ansla ion? A comp e-
hensi e e alua ion.h ps://a xi .o g/abs/2302.09210.
House, Juliane. 1981. A model o ansla ion quali y assessmen . 2nd edn. Tübin-
gen, Ge many: Gun e Na .
ITI. 2024. Slow ansla ion mani es o. h ps://www.i i.o g.uk/disco e /policy/
slow- ansla ion-mani es o.h ml.
Japan Associa ion o T ansla o s. 2024. S a emen on he public and p i a e sec o
ini ia i e o use AI o high- olume ansla ion and expo o manga. h ps://
p imes.jp/main/h ml/ d/p/000000001.000143535.h ml.
Jiao, Wenxiang, Wenxuan Wang, Jen- se Huang, Xing Wang, Shuming Shi &
Zhaopeng Tu. 2023. Is Cha GPT a good ansla o ? Yes wi h GPT-4 as he engine.
h ps://a xi .o g/abs/2301.08745.
LITHME P ojec . 2021. Language in he human-machine e a. h ps://li hme.eu/.
Miya a, Rei, Masa u Yamada & Kyo Kageu a. 2023. Me alanguages o dissec ing
ansla ion p ocesses: Theo e ical de elopmen and p ac ical applica ions. Lon-
don, UK: Rou ledge. h ps : / / www . ou ledge . com / Me alanguages - o -
Dissec ing-T ansla ion-P ocesses-Theo e ical-De elopmen -and-P ac ical-
Applica ions/Miya a-Yamada-Kageu a/p/book/9781032168951.
Nida, Eugene A. & Cha les R. Tabe . 1969/2003. The heo y and p ac ice o ans-
la ion. Leiden, Ne he lands: B ill.
Papineni, Kisho e, Salim Roukos, Todd Wa d & Wei-Jing Zhu. 2002. BLEU: A
me hod o au oma ic e alua ion o machine ansla ion. In P oceedings o
he 40 h Annual Mee ing on Associa ion o Compu a ional Linguis ics, 311–318.
Philadelphia, USA: Associa ion o Compu a ional Linguis ics. DOI: 10.3115/
1073083.1073135. h ps://www.aclweb.o g/an hology/P02-1040.
Rei, Rica do, C aig S ewa , Ana C. Fa inha & Alon La ie. 2020. COMET: A neu-
al amewo k o m e alua ion. In P oceedings o he 2020 Con e ence on Em-
pi ical Me hods in Na u al Language P ocessing (EMNLP 2020), 2685–2702. On-
line: Associa ion o Compu a ional Linguis ics. DOI: 10.18653/ 1/2020.emnlp-
main.213. h ps://www.aclweb.o g/an hology/2020.emnlp-main.213.
Saye s, Da e, Rui Sousa-Sil a, S ia lana Höhn, Lule Ahmedi, Kais Allki i-
Me soja, Dimi a Anas asiou, Š e an Beňuš, Lynne Bowke , Elio By yçi, Ale-
jand o Ca ala, Anila Çepani, Rubén Chacón-Bel án, Sami Dadi, Fisnik Dalipi,
Vladimi Despo o ic, Agnieszka Doczekalska, Sebas ian D ude, Ka ën Fo ,
Robe Fuchs, Ch is ian Galinski, Fede ico Gobbo, Tunga Gungo , Siwen Guo,
Klaus Höckne , Pe alea Láncos, Tome Libal, Tommi Jan unen, Dewi Jones,
Blanka Klimo a, Emine kan Ko kmaz, Sepesy Maučec Mi jam, Miguel Melo,
103
Masa u Yamada
Fanny Meunie , Be ina Migge, Ba bu Mi i elu Ve ginica, Au élie Né éol,
A ianna Rossi, An onio Pa eja-Lo a, Ch is ina Sanchez-S ockhamme , Aysel
Şahin, Angela Sol an, Claudia So ia, Sa ang Shaikh, Ma co Tu chi & Sule
Yildi im Yayilgan. 2021. The dawn o he human-machine e a: A o ecas o new
and eme ging language echnologies. h ps://hal.science/hal-03230287.
Sla o . 2024. F ench ansla o s socie y akes ough s ance on AI ansla ion. h ps://
sla o .com/ ench- ansla o s-socie y- akes- ough-s ance-on-ai- ansla ion-
genai/.
Ti kkonen-Condi , Sonja. 2005. The moni o model e isi ed: E idence om p o-
cess esea ch. Me a 50(2). 405–414. DOI: 10.7202/010990a .
Venu i, Law ence. 1995. The ansla o ’s in isibili y: A his o y o ansla ion. Lon-
don, UK: Rou ledge.
Vinay, Jean-Paul & Jean Da belne . 1958/1995. Compa a i e s ylis ics o F ench
and English: A me hodology o ansla ion. Ams e dam, Ne he lands: John Ben-
jamins. DOI: 10.1075/b l.11.
Wang, Longyue, Chenyang Lyu, Tianbo Ji, Zhi ui Zhang, Dian Yu, Shuming Shi &
Zhaopeng Tu. 2023. Documen -le el machine ansla ion wi h la ge language
models. In Houda Bouamo , Juan Pino & Kalika Bali (eds.), P oceedings o he
2023 Con e ence on Empi ical Me hods in Na u al Language P ocessing, 16646–
16661. Singapo e: Associa ion o Compu a ional Linguis ics. DOI: 10.18653/ 1/
2023.emnlp-main.1036.
Way, Andy. 2025. Wha does he u u e hold o ansla ion echnologies? In S e-
an Baumga en & Michael Tiebe (eds.), The Rou ledge handbook o ansla ion
echnology and socie y, 448–461. London, UK: Rou ledge.
Yamada, Masa u. 2023. Op imizing machine ansla ion h ough p omp engi-
nee ing: An in es iga ion in o Cha GPT’s cus omizabili y. In Masa u Yamada
& Félix do Ca mo (eds.), P oceedings o Machine T ansla ion Summi XIX, Vol.
2: Use s T ack, 195–204. Macau, China: Asia-Paci ic Associa ion o Machine
T ansla ion. h ps://aclan hology.o g/2023.m summi -use s.19.
Yang, Zhishen, Tosho Hi asawa, Edison Ma ese-Taylo & Naoaki Okazaki. 2024.
La ge language models as manga ansla o s: A case s udy. In P oceedings o
he 30 h Annual Mee ing o he Associa ion o Na u al Language P ocessing,
2012–2017. Tokyo, Japan: The Associa ion o Na u al Language P ocessing.
h ps://www.anlp.jp/p oceedings/annual_mee ing/2024/pd _di /P7-13.pd .
Zhang, Biao, Ba y Haddow & Alexand a Bi ch. 2023. P omp ing la ge language
model o machine ansla ion: A case s udy. In P oceedings o he 40 h In e na-
ional Con e ence on Machine Lea ning (ICML’23). Honolulu, USA: JMLR.o g.
104
