Pa allel Co po a Spanish (PaCo ES): A collec ion o mul i unc ional pa allel
co po a
Co pus pa alelos del español (PaCo ES: una colección de co pus pa alelos
mul i uncionales
IRENE DOVAL REIXA, Uni e si y o San iago de Compos ela (Spain)
MARÍA TERESA SÁNCHEZ NIETO, Uni e si y o Valladolid (Spain)*
Abs ac
This pape in oduces PaCo ES, a compila ion o pa allel bidi ec ional co po a wi h
Spanish as he ocal language. I is cu en ly comp ised o ou co po a in
Ge man/Spanish, English/Spanish, F ench/Spanish, and Chinese/Spanish, which a e
accessible online wi hou es ic ions.
The aim o his pape is o p o ide adequa e documen a ion o he PaCo ES p ojec
and i s con en s o he co pus esea ch communi y, while unde sco ing i s
mul i unc ionali y. Fi s , gaps a e iden i ied in exis ing pa allel bi- and mul ilingual
co po a ea u ing Spanish as one o hei componen s. Then, he c i e ia unde lying he
design and a chi ec u e o he PaCo ES co po a a e ou lined. The main sec ion o he
pape is de o ed o he di e en s eps in he compila ion o he co po a, wi h special
emphasis on he alignmen p ocess and i s e iew. Finally, a en ion is di ec ed owa ds
he h ee sea ch le els a ailable o que ying he co po a and he display ea u es. This
ende s PaCo ES co po a a aluable esou ce o a wide ange o applica ions and use
g oups ac oss di e se linguis ic ields, as well as language eache s, p ac i ione s, and
language lea ne s alike.
Key wo ds: pa allel co po a, bidi ec ional co po a, co pus mul i unc ionali y, co pus
applica ions, co pus compila ion, co pus alignmen
• NOTE: Bo h au ho s o his a icle ha e con ibu ed equally o i s c ea ion, and
he o de o au ho ship ollows only an alphabe ical c i e ion.
Resumen
El obje i o de es e abajo es p opo ciona a los in es igado es en el ámbi o de la
lingüís ica de co pus documen ación adecuada del P oyec o PaCo ES y los ecu sos que
comp ende, haciendo especial én asis en la mul i uncionalidad de es os. En p ime
luga , se iden i ican las lagunas exis en es en el pano ama de los co pus pa alelos
bilingües y mul ilingües que incluyen el español como una de sus lenguas. A
con inuación, se exponen los c i e ios que han guiado el diseño y la a qui ec u a de los
co pus de PaCo ES. La sección p incipal del abajo se dedica a las ases de la
compilación de los co pus, con especial a ención al p oceso de alineado y de su
e isión. Finalmen e, se sub aya la exis encia de es ni eles de búsqueda disponibles
pa a in e oga el co pus y se explica las uncionalidades de la in e az y de la
p esen ación de esul ados. Es as úl imas ca ac e ís icas hacen de los co pus de
PaCo ES un ecu so alioso pa a un amplio ango de aplicaciones y ipos de usua ios:
in es igado es de di e en es ámbi os de la lingüís ica, docen es de lenguas, usua ios
p o esionales de la lengua o ap endices.
Palab as cla e: co pus pa alelos, co pus bidi eccionales, mul i uncionalidad de los
co pus, aplicaciones de los co pus, compilación de co pus, alineamien o de co pus
1 In oduc ion
The compila ion o ex ensi e pa allel co po a has been he key ac o in he
spec acula de elopmen o s a is ical (and mo e ecen ly neu al) machine ansla ion
sys ems and in o he na u al language applica ions, such as speech ecogni ion, speech-
o- ex and ex - o speech echnologies, au oma ic documen summa izing, and many
o he s. On he o he hand, pa allel co po a ha e been also used o in es iga ing
linguis ic and communica i e phenomena, mos no ably languages in con as and
ansla ion. Fu he mo e, pa allel co po a can ind uses in applied a eas such as
lexicog aphy, language and ansla ion eaching and lea ning. O all he scena ios, he
la e (i.e., ansla ion and language lea ning and eaching wi h pa allel co po a) elies
he mos on he a ailabili y and usabili y o pa allel esou ces.
PaCo ES is a collec ion o bilingual co po a wi h Spanish as he cen al language
ha can be easily and eely que ied online. The oo s o he PaCo ES p ojec da e back
o 2015 wi h he c ea ion o he Ge man<>Spanish Pa allel Co pus (PaGeS)
1
(Do al e
al., 2019, Do al & Jiménez Juliá, 2019, Do al & Sánchez Nie o, 2022).
A he ime o w i ing, he English<>Spanish Pa allel Co pus (PaEnS) (Au ho ,
yea ) and he F ench<>Spanish Pa allel Co pus (PaF eS) ha e al eady been launched
and can be accessed online. Mo eo e , he Chinese<>Spanish Pa allel Co pus ( ) is
cu en ly being compiled and will be accessible soon. All hese co po a d aw on he
same a chi ec u e and philosophy as he ini ial PaGeS co pus.
Conside ing ha he c ea ion o pa allel co po a is a demanding ask ha consumes
signi ican esou ces, ou aim om he beginning was o design he PaCo ES co po a o
be mul i unc ional, capable o se ing a wide ange o applica ions and use g oups,
such as language lea ne s and eache s, ansla o s, ansla o aine s and educa o s,
ansla o ainees and s uden s, language, communica ion and ansla ion schola s, and
1
Fo he co po a’s URLs see oo no e 8.
lexicog aphe s. I goes wi hou saying ha hese co po a can be used as a eliable sou ce
o ex s o aining machine ansla ion sys ems.
The pu pose o his pape is o p o ide adequa e documen a ion o he PaCo ES
p ojec and i s con en s o he co pus esea ch communi y. In sec ion 2, PaCo ES is
mapped on o he landscape o hose pa allel bi- and mul ilingual co po a ha a e in one
way o ano he compa able o i s own. In Sec ion 3, he c i e ia leading o he design
and he make-up o he PaCo ES co po a a e explained. The asks conduc ed be o e
alignmen a e de ailed in Sec ion 4, while Sec ion 5 del es in o he alignmen p ocess.
Sec ion 6 in oduces he ea u es o he co pus sea ch engine unde lying all PaCo ES
co po a, o e ing insigh s in o he co pus que y sys em and esul s display. Finally,
Sec ion 7 w aps up he main conclusions o he pape .
2 Rela ed wo k
In his sec ion, we will p o ide a b ie o e iew o a selec ion o pa allel co po a
ha a e compa able o he PaCo Es collec ion wi h he aim o disco e ing whe he hese
esou ces a e mul i unc ional as de ined o he PaCo ES co po a in Sec ion 1.
The c i e ia o he inclusion o he esou ces in his o e iew a e (i) being eely
accessible and (ii) ha ing Spanish as one o he included languages. As o co pus
accessibili y, a dis inc ion can be made be ween co po a ha can be accessed and
que ied online h ough a web-based co pus in e ace on he one hand, and co po a
which a e a ailable o download online o u he p ocessing and elabo a ion on he
o he . As o he p esence o he Spanish language in he selec ed co po a, a en ion
mus be paid o (a) whe he he Spanish ex s included a e he esul o ei he di ec o
indi ec ansla ion p ocesses (i.e., as a ge ex s ansla ed om ano he language and
subsequen ly pa allelized o exis ing ansla ions o he same ex in o u he
language/s); and o (b) whe he he s a us o Spanish as ei he he sou ce o a ge
language is known.
Fou selec ed pa allel esou ces ha mee he a o emen ioned c i e ia a e b ie ly
p esen ed below.
The g owing collec ion o bilingual and mul ilingual pa allel co po a included in
OPUS
2
(Tiedemann, 2012) o e s no only he bigges collec ion o online sea chable
pa allel esou ces, bu also ools o uploading co po a o he collec ion and p e ained
machine ansla ion models. A numbe o he esou ces in OPUS a e PoS- agged
(Tiedemann 2012, pp. 2216) and can be sea ched online wi h he help o mul ilingual
conco dance ools based on he que y language used in he Co pus Wo kbench co pus
manage (Tiedemann, 2012). Ex e nal use s can con ibu e hei own pa allel esou ces
o he co po a collec ion (Aulamo and Tiedemann, 2019, p.1). A he ime o w i ing, 81
esou ces can be sea ched and downloaded o he Ge man/Spanish language pai , 122
o English/Spanish, 93 o F ench/Spanish and 19 o Chinese/Spanish.
Mul ilingwis
3
is one o he mos success ul a emp s o make a pa allel esou ce
a ailable h ough an online sea ch in e ace. The sys em allows he use o que y he
pa allel co pus o he Eu opean Pa liamen “ o spo ing ansla ions o con en wo ds
and mul i-wo d uni s” (Mul ilingwis, 2023). The co pus con ains 220 million wo ds and
wo d alignmen s o abou 40 million wo ds. The suppo ed languages a e English,
Spanish, Ge man, F ench, and I alian.
In e Co p,
4
a pa o he Czech Na ional Co pus, is a mul ilingual pa allel co pus
whose 8 h elease includes abou 1.4 billion wo ds in 38 languages and 174 million
wo ds in Czech (Rosen, 2016, p. 22). I is comp ised o a co e co pus o mainly li e a y
ex s and collec ions o o he ex s, which a e ob ained om o he mul ilingual, eely
a ailable esou ces. The ex s a e au oma ically aligned a he sen ence le el ia Czech,
which is he pi o language. The ex s a e also lemma ized and pa ly agged (Rosen,
2016, p. 23) and can be eely que ied on he in e ne .
Linguee
5
is an online mul ilingual dic iona y combined wi h an engine ha looks up
wo ds and wo d g oups in billions o ansla ed ex s and e u ns aligned bilingual
2
See all he OPUS esou ces a h ps://opus.nlpl.eu/ and he OPUS sea ch in e ace a
h ps://opus.nlpl.eu/bin/opuscqp.pl.26.
3
h ps://pub.cl.uzh.ch/p ojec s/spa cling/mul ilingwis/ and
h ps://pub.cl.uzh.ch/p ojec s/spa cling/mul ilingwis2.demo/.
4
h ps://kon ex .ko pus.cz.
5
h ps://www.linguee.de/.
examples. Equi alences can cu en ly be gene a ed in mo e han 20 languages (Hein,
2022, p. 325), including, o cou se, English, Ge man, Spanish, Chinese, and F ench.
Among he mul ilingual pages c awled by Linguee's mul ilingual co pus compila ion
sys em a e also si es ha ely on unsupe ised machine ansla ion (Alonso, 2013, pp.
22-23).
In summa y, while we can conclude ha all ou o he a o emen ioned pa allel
esou ces undoub edly ha e in insic alue, hey also bea ce ain disad an ages o bo h
esea ch in con as i e linguis ics and ansla ion, and o o eign language eaching and
lea ning and hus canno be conside ed mul i unc ional as desc ibed in sec ion 1. Mos
o hese esou ces a e ei he limi ed o speci ic ex ual a ie ies (mainly legal,
adminis a i e, o echnical language), o nei he he o iginal language no he
ansla ion p ocess ca ied ou can be p ecisely de e mined. Finally, in some cases, he
sea ch and display ea u es a e no adap ed o di e en use g oups.
Mul ilingwis and many o he OPUS pa allel co po a ea u e o a la ge ex en
examples o specialized communica ion. Business, adminis a i e, and legal language,
as well as medical o echnical language, show li le lexical and mo phosyn ac ic
a ia ion and end owa ds o mulaic language use. These domains a e hus unsui able
o eaching/lea ning o non-speci ic pu poses, as he le el o di icul y o he
s uc u es he ein is oo high. A he same ime, Linguee p o ides many pa allel ex s
o a numbe o language pai s. Howe e , along wi h he linguis ic esou ces p oduced
by in e na ional ins i u ions, such as he Eu opean Commision, Linguee also makes
o he ma e ials a ailable which ha e no been p oduced o e ised by a human
ansla o and he sou ce language is no known, some hing ha holds ue o a numbe
o he pa allel esou ces in OPUS as well. Tex s sc aped om un e i ied websi es ia
au oma ic ex ac ion and alignmen a e likely o con ain e o s and misalignmen s. The
In e Co p co po a ha e been compiled wi h Czech as he pi o language, so ha he
English/Spanish, Ge man/Spanish e c. alignmen s a e indi ec ansla ions; hese a e
e y use ul in ce ain eaching and esea ch con ex s, bu o mo e limi ed use in o he s.
In he case o OPUS, he sea ch language (CQL) in he mul ilingual sea ch in e ace
6
equi es p io aining, which may discou age non-expe use s ha y on- he- ly
6
h ps://opus.nlpl.eu/bin/opuscqp.pl.
que ies on hei own o he i s ime. The same migh apply as well o ce ain non-
expe use s app oaching he KonTex in e ace o que y In e Co p.
7
In summa y, i can be said ha he PaCo ES co po a ha e come o ill a gap wi hin
he landscape o bilingual co po a o he Spanish-speaking wo ld. As he ollowing
sec ions will show, he s ong ocus on he quali y o he con en , he conside able size
o he co po a (sec ion 3) and hei usabili y (see sec ion 6) makes hem amenable o a
b oad scope o po en ial use s.
3 Composi ion o he PaCo ES co po a
In his sec ion we will desc ibe in de ail he di e en ypes o da a included in he
PaCo Es co po a. The exis ing co po a o da e a e as ollows, in o de o c ea ion:
− The Ge man<>Spanish Pa allel Co pus, PaGeS
− The English<>Spanish Pa allel Co pus, PaEnS
− The F ench<>Spanish Pa allel Co pus, PaF eS, and
− The Chinese<>Spanish Pa allel Co pus, PaCheS
8
Gi en ha his is an ongoing p ojec ha began se e al yea s ago, he di e en
bilingual co po a cu en ly ind hemsel es a a ious s ages o comple ion. As a esul ,
hey di e no only in size bu also empo a ily in e ms o he ypes o indexed ex s. In
he case o he Chinese<>Spanish and F ench<>Spanish pa allel co po a, he co e
co po a (see 3.1) a e no ye online since he numbe o ex s ha ha e unde gone
e iew is s ill oo small. Ne e heless, he gene al guidelines o compiling he di e en
co po a emain consis en ac oss all ou .
In PaCo Es he e a e wo la ge g oups o well-di e en ia ed da a: he i s consis s
o he co e co pus, and he second is made up o he so-called “supplemen s”. Bo h
g oups di e in hei ea u es, he p ocessing p ocedu es hey unde go, and he ex en o
which e iews a e conduc ed du ing he a ious p ocessing s eps.
7
h ps://www.ko pus.cz/kon ex /que y?co pname=syn2020.
8
The URLs in he o de lis ed: www.co puspages.eu, www.co puspaens.eu, www.co puspa es.eu,
www.co puspaches.eu.
3.1 PaCo ES co e co po a
Taking in o conside a ion he pu pose o he co po a and he di e en applica ions
o which hey a e in ended, we ha e d awn up a se o guidelines o he da a collec ion
o he co e co pus as explained in sec ions 1 and 2. These include he ollowing
ea u es:
− bo h he o iginal da a and he ansla ions mus ha e unde gone e i iable quali y
con ols
− he e mus be an unambiguous iden i ica ion o he o iginal and he ansla ed
language
− he ansla ions mus ha e been ca ied ou di ec ly -wi hou a pi o language- om
he o iginal by human ansla o s
− he ex s should be as close as possible o he cu en s anda d a ie y o he
language, and belong o a gene al, non-specialized egis e .
In addi ion o hese c i e ia, i should be no ed ha compiling a pa allel co pus, in
con as o a monolingual one, p esen s signi ican cons ain s in e ms o da a
a ailabili y as he as majo i y o ex s a e no ansla ed (Zane in, 2012, p. 41).
3.1.1 Co pus Design and Resea ch Po en ial
This sec ion will de ail he design o he co pus and conside i s po en ial as a
esea ch ool.
In o de o he ex s o mee he c i e ia ou lined abo e, we ha e chosen o limi ou
scope o ex s om epu able publishe s, he e o e ensu ing ha bo h he o iginals and
he ansla ions ha e unde gone igo ous quali y con ols (Do al, 2018).
Each co e co pus consis s o wo se s o ex s:
− o iginal ex s in he o eign language (FL)
9
and hei ansla ions in o Spanish
− o iginal ex s in Spanish and hei ansla ions in o he o eign language.
Bo h se s a e simila in size and he ypes o ex s hey con ain. As a esul , all he
co e co po a a e bidi ec ional a nd balanced in e ms o ansla ion di ec ion.
9
Wi h he e m ' o eign language' (FL) we e e o he ou languages included so a in he Paco ES
collec ion: Ge man, English, F ench, and Chinese.
Rega ding he la e , i should be no ed ha due o he egula inco po a ion o new
wo ks, sligh disc epancies may occasionally occu .
This bidi ec ional and balanced design o he co e co po a enables a b oade ange
o analyses and applica ions compa ed o unidi ec ional co po a. (c . S ieg Johansson,
2004, p. 61).
This design acili a es he compa ison o o iginal ex s and ansla ions ac oss
languages (Figu e 1) and allows o bilingual analyses om bo h he FL o Spanish and
Spanish o FL. This app oach is essen ial o s udying how pa e ns a y ac oss
languages. Bidi ec ionali y is pa icula ly c ucial because, in many cases, ansla ion
equi alen s a e no biuni ocal, i.e., he ansla ion o i em A (sou ce ex ) in o B
( ansla ion) does no necessa ily imply ha B (sou ce) is ansla ed back in o A
( ansla ion). A bidi ec ional co pus allows o he measu emen o he pe cen age o
mu ual co espondence (MC) be ween wo speci ic i ems, anging om 0% o no
co espondence o 100% o comple e co espondence. (Beng Al enbe g, 1999, p. 254).
Figu e 1: PaCo ES as pa allel co po a: o iginal ex s and ansla ions FL and Spanish
Mo eo e , bidi ec ional co po a also enable compa isons be ween o iginal and
ansla ed ex s wi hin each language (Figu e 2) and be ween ansla ed ex s ac oss
languages (Figu e 3). Hence, he co po a can be used o esea ch in o speci ic
ansla ion phenomena such as ansla ionese o ansla ion uni e sals, i.e., ecu ing
linguis ic pa e ns o ea u es obse ed in ansla ed ex s ac oss di e en languages
(Mona Bake 1993, pp. 742-746).
Figu e 2: PaCo ES as pa allel co po a: o iginal ex s and ansla ions in each language
Figu e 3: PaCo ES as pa allel co po a: ansla ions ac oss languages
FL
o iginal Spanish
ansla ion Spanish
o iginal FL
ansla ion
FL
o iginal FL
ansla ion Spanish
o iginal Spanish
ansla ion
FL
ansla ion Spanish
ansla ion
All hese no maliza ion asks a e execu ed wi h he help o egula exp essions. The
pieces o so wa e used o his pu pose a e ei he he ee so wa e No epad++ o he
licensed so wa e EmEdi o .
19
As o he s uc u al anno a ion, pa s and chap e s o a wo k a e signaled wi h
speci ic ags. The ags o he pa s o a wo k a e placed be ween cu ly b acke s ({}) and
ags o chap e s a e placed be ween angula b acke s (<>); a abic numbe s a e used.
These ags a e use ul a a la e s age o he p ocessing: du ing he alignmen p ocess, he
aligning algo i hm will use hem as ancho s o ecognizing he beginning and he end
o pa allel ex sec ions.
The ex alinguis ic anno a ion o ex me ada a a e collec ed in a s uc u ed manne
ia a sp eadshee ha allows o a con e sion o he ele an in o ma ion in o xml
o ma . The collec ed me ada a can be di ided in o wo di e en g oups: ex -le el
me ada a, and p ojec -le el me ada a.
Tex -le el me ada a include, i s ly, he na u e o he ex (o iginal o ansla ion),
he s a us o i s Spanish e sion (ei he o iginal o ansla ion), and language o he
o iginal e sion. These me ada a esul in a i e-numbe code iden i ying (i) he co pus
o which he ex belongs (PaGeS, PaEnS, PaChEs, PaF eS), (ii) he ansla ion
di ec ion, and (iii) he ex i sel . Fu he ex -le el me ada a a e he ex ’s dialec al
a ian ( esul ing in a ou -le e code, e.g., eses o Eu opean Spanish, esch o Chilean
Spanish, e c.),
20
i s gen e/subgen e, and e iewe ( eam membe [s] in cha ge o he
p ep ocessing asks o ha ex and he manual e ision o he alignmen ).
P ojec -le el me ada a include in o ma ion abou he ex ma e ial ha esul s om
he p ocessing asks: cha ac e , oken, and wo d numbe o he o eign language and
o he Spanish ex .
5 Segmen a ion and alignmen
5.1 Segmen a ion o he da a
19
h ps://no epad-plus-plus.o g/ and h ps://www.emedi o .com/.
20
See he ull lis a h ps://www.co puspages.eu/co pus/help/help?lang=en .
Segmen a ion in ol es b eaking up a s ing o w i en ex in o smalle p ocessing
uni s. The ype o uni s (documen , pa ag aph, sen ence, o wo d) de e mines he ype o
alignmen ha can be subsequen ly es ablished. P ope segmen a ion is essen ial o he
pe o mance o he au oma ic alignmen p ocess since e oneous segmen a ion o en
leads o a numbe o misalignmen s, as Jö g Tiedemann (2011, p. 9) indica es.
As in mos pa allel co po a, in he Paco ES p ojec he basic uni o alignmen is he
sen ence, which is desc ibed in de ail below. The sen ence segmen a ion ask in ol es
iden i ying sen ence bounda ies be ween di e en sen ences (Palme , 2000, p. 11). The
c i e ia o a sen ence bounda y a e a punc ua ion ma k such as a pe iod, exclama ion
poin , ques ion ma k (.!?), o occasionally a semicolon, ollowed by whi espace. Bu
his ask poses a ew challenges, p ima ily because hese ma ks a e no unambiguous.
So, a pe iod may deno e in Spanish, like in many o he Eu opean languages, in addi ion
o he end o a sen ence, an abb e ia ion, an ellipsis, o a decimal poin , among o he s.
To p e en he inco ec assignmen o sen ence bounda ies, he inal punc ua ion ules
accommoda e hose excep ions.
Ano he majo issue in he segmen a ion p ocess is ha in li e a y ex s, ce ain
au ho s o wo ks, due o hei speci ic li e a y s yle,
21
may ha e long s e ches o ex
wi hou any inal punc ua ion ma ks. In his case, au oma ic segmen a ion leads o e y
long segmen s ha would be unsui able o he pu poses o he p ojec . The sho e he
bisegmen s he sea ch engine e u ns, he easie i is o he use o spo equi alences o
a gi en phenomenon. Fo his eason, bo h segmen s a e manually subdi ided a e
au oma ic alignmen . The maximum numbe o cha ac e s allowed o a segmen in
ei he language has been se o 300. Segmen s exceeding his leng h a e spli by
inse ing manual b eak ma ks in sui able places in bo h he L1 and he L2. The a e age
segmen leng h (ASL) o a wo k is calcula ed by di iding he o al numbe o cha ac e s
by he numbe o bisegmen s and mul iplying by 1000.
𝐴𝑆𝐿 = 𝐶ℎ𝑁(𝐿1)+ 𝐶ℎ𝑁(𝐿2)
𝐵𝑖𝑠𝑒𝑔𝑁 ∗ 1000
21
Fo example, au ho s like Ja ie Ma ías o Leona do Padu a s and ou o w i ing long pieces o ex
wi hou inal punc ua ion ma ks.
The a e age segmen leng h (ASL) anges om 0 o 1, wi h a lowe sco e indica ing
ha ewe segmen s will need o be manually cu . The ASL can a y signi ican ly om
wo k o wo k depending on he au ho 's s yle.
22
The a e age ASL o all ex s in he
co e co pus a e subsegmen a ion is 0.161. In he o e all manual e iew, his ask is he
mos ime-consuming.
5.2 Alignmen and manual e iew
Sen ence alignmen is he ask o iden i ying co espondences be ween sen ences in
each hal o a bi ex . The i s p oposals o au oma ic sen ence alignmen da e back o
he ea ly 1990s and can be ca ego ized in o leng h-based and lexical ma ching
app oaches (Tiedemann, 2011, p. 38). To add ess he limi a ions o hese me hods,
hyb id app oaches such as Hunalign
23
(Va ga e al., 2007) and Ga gan ua (B aune and
F ase , 2010) we e la e in oduced, combining bo h leng h-based and lexical ma ching
s a egies.
In he Paco ES p ojec he hyb id open-sou ce so wa e LF-Aligne , ha elies on
he Hunalign algo i hm, and occasionally Ga gan ua
24
a e used because hey ha e
achie ed he bes esul s in se e al es s.
25
LF-Aligne p oduces no only one- o-one
alignmen s, bu many- o-one and one- o-many alignmen pai ings as well, which a e
needed o ensu e ha all he inpu sen ences a e ac ually aligned. (Xu e al., 2015, p. 5).
22
To men ion wo ex eme examples, Leona do Padu a's wo k The Man Who Lo ed Dogs (11048) has an
ASL o 0.556, while Louise Penny Louise’s The Beau i ul Mis e y (10062) has an ASL o 0.11.
23
This ool is ep esen a i e o he cu en s a e-o - he a in sen ence alignmen and has been widely used
in mul ilingual p ojec s, like Eu opa l (h ps://www.s a m .o g/eu opa l/), Opus co po a
(h ps://opus.nlpl.eu/ ools.php), In e Co p (h ps://in e co p.ko pus.cz/?lang=en), e c.
24
LF-Aligne : h ps://sou ce o ge.ne /p ojec s/aligne /,
Ga gan ua: h ps://gi hub.com/b aune e/Ga gan ua.
25
The ollowing aligne s we e es ed, among o he s: ABBYY Aligne 2.0 (h ps://abbyy.s o e/aligne -
co po a e/), Youalign (h ps://youalign.com/) and WinAlign
(h ps://www. ados.com/solu ions/ ansla ion-alignmen /), comme cial ools o aligning pa allel ex s
and c ea ing T ansla ion Memo y da abases; Champollion h p://sou ce o ge.ne /p ojec s/champollion/,
a lexicon-based sen ence aligne ; bi ex h ps://sou ce o ge.ne /p ojec s/bi ex 2 mx/.
The ou pu is a sequence o bilingual sen ence pai s (bisegmen s) in ab delimi ed
o ma .
The accu acy o alignmen depends la gely on he ype o ex s. I is well known ha
all sys ems pe o med poo ly on li e a y ex s compa ed o o he ypes o ex s such as
adminis a i e o echnical (Xu e al., 2015, p. 1; Zane in, 2012, p. 155). Mo eo e ,
wi hin ic ional ex s, he deg ee o co espondence o he bi ex a ies depending on he
au ho , he ansla o , he ex s hemsel es, and he di ec ion o he ansla ion.
26
A e he ex has been au oma ically aligned, each bisegmen ecei es a unique ID,
composed o he 5 numbe s o he wo k (see abo e) and 7 sequen ial numbe s,
sepa a ed by a hyphen, which co espond o he bisegmen i sel . This ID is c ucial o
he manual e iew o he alignmen , as i allows he segmen s o be so ed wi hou
losing he o iginal o de o he ex and ecalled h oughou he di e en s eps ha will
be desc ibed in he subsequen pa ag aphs o his subsec ion.
A e subsegmen a ion (see Sec ion 5.1), he mapping be ween he co esponding
sen ences is checked. The i s s ep is o check o au oma ic co espondences o 1:0 o
0:1. These alignmen s may be due o omissions o addi ions in he a ge ex (L2) wi h
espec o he sou ce ex (L1). In bo h cases, he ollowing is indica ed: i a piece o ex
is omi ed in L2, i is ma ked as [n_ _s] (non- ansla ed segmen ); i i has been added in
L1, i is ma ked as [a_s_ ] ( ex added in he ansla ion). This can be seen in Table 6.
These ma ks, which migh seem unin e es ing o some use s, play a c ucial ole in
analyzing speci ic ins ances o ansla o in e en ion (Munday, 2007).
Table 6: Omission and addi ion in he ansla ion (L2)
01076-0000299
Ponía eno mes can idades de
g asa de oca en odo, al es ilo
de su Belg ado na al,
Sie benu z e iesige Mengen
Gänseschmalz, wie sie es aus ih e
se bischen Heima kann e,
26
Pa icula ly challenging is he alignmen o Chinese li e a y ex s. The abo e-men ioned aligne s ha e
he e a much lowe le el o accu acy. Since i equi es e y labo ious and ime-consuming co ec ion
wo k, he e a e e y ew ex s in he co e co pus o PaCheS, so i is no o e ed online o he ime
being. To go deepe in o his issue would go a beyond he scope o his pape . We simply men ion
ha we a e es ing wi h non-mono onic so wa e, speci ically wi h Vecalign (Thompson /Koehn 2019)
I employs sen ence embeddings o es ima e alignmen and i has shown p omising esul s in aligning
sen ences wi h challenging cases whe e he sen ence s uc u e di e s signi ican ly, as in he case o
Chinese.
01076-0000300
y yo ya había engo dado un
pa de kilos.
[n_ _s]
11041-0002935
—¿No la is e, anoche?
“Didn’ you see he las nigh ?”
11041-0002936
[a_s_ ]
Osca asked.
11041-0002937
—¡Cla o que la i!
“O cou se I saw he !”
In o he cases, alignmen s o 1:0 and 0:1 a e due o misalignmen s esul ing om a
sen ence ha ing been spli in o mo e han one sen ence in he a ge ex
(co espondence 1:<1, Table 7) o , in e sely, se e al sen ences ha e been me ged in o
one (co espondence <1:1, Table 8). He e he necessa y co ec ions a e made.
Table 7: Inco ec alignmen 1:2
11048-
0006152
—¿Te a a da una abie a? —la oz
de Tom des ilaba i onía—.
“A e you going o ha e a an um?”
11048-
0006153
No e oy a epe i lo que ya sabes.
Tom’s one was sa cas ic.
11048-
0006154
“I’m no going o epea wha you
al eady know.
Table 8: Inco ec alignmen 2:1
11048-
0002714
—Mi amigo mu ió…
11048-
0002715
Y cuando yo mue a, y cuando mue a
la o a única pe sona que, según sé,
conoce casi odos los de alles, esa
his o ia se pe de á.
My iend died… and when I die, and
when he only o he pe son dies who, as
a as I know, is amilia wi h all he
de ails, ha s o y will be los .
Finally, hose unbalanced bisegmen s a e e iewed ha a e mo e likely o be
misaligned in e ms o he numbe o cha ac e s. Since we ha e ound ha he ele ance
o he imbalance depends on he numbe o cha ac e s, he less cha ac e s he bisegmen
has he less ele an he imbalance is. To measu e he le el o imbalance (IL) we apply
he ollowing a io, whe e x ep esen s he o al numbe o cha ac e s in he L1 segmen
and y in he L2 segmen . The o mula uses loga i hms o educe he e ec o he numbe
o cha ac e s.
𝐼𝐿 = 1 𝑙𝑜𝑔(𝑥)− 𝑙𝑜𝑔(𝑦)
𝑚𝑎𝑥(𝑙𝑜𝑔(𝑥),𝑙𝑜𝑔(𝑦))
The close he alues a e o 1, he mo e balanced he wo segmen s a e and he mo e
unlikely a misalignmen is. In he manual e iew we ocus on he segmen s wi h a alue
below <0.5.
Table 9: Unbalanced inco ec alignmen
00081-
0004868
Bajó las en anillas y condujo
así, suspi ando con usa y
salpicando el in e io del coche
con las go as que lle aba
adhe idas po ue a.
Ve s ö seu z e sie au und ließ die
Fens e scheiben he un e .
0,36
00081-
0004869
El sonido del elé ono, que
eposaba en el asien o del
copilo o, in e umpió un hilo de
pensamien os oscu os.
Die T op en, die außen da an
geha e ha en, sp i z en ins
Wageninne e. In diesem Augenblick
klingel e das Handy, das au dem
Bei ah e si z lag.
0,42
No e ha his a io can gi e bo h alse posi i es and alse nega i es. Table 10
p esen s examples o alse posi i es, i.e., unbalanced segmen s ha ne e heless a e
co ec ly aligned.
Table 10: False posi i e co ec alignmen
01076-0000764
Un escalo ío me a enazó la
espalda.
Mich schaude e.
0,29
10043-0010055
Fa om i .
Las cosas no e an así, ni mucho
menos.
0,21
Occasionally sen ences a e eo de ed when ansla ing, bu Hunalign, like mos
sen ence aligne s, is mono onic, meaning ha he ela i e o de o sen ences is
p ese ed and i is unable o come up wi h c ossing alignmen s, i.e., segmen s A and B
in one language co esponding o segmen s B' A' in he o he language. In hese cases, a
simple ope a ion o me ging he sen ences in o la ge uni s mus be pe o med o ind a
mono onic mapping o he co esponding pa s, e en a he cos o longe segmen s, as
in he ollowing example:
Table 11: C ossing alignmen
11048-
000283
Demasiado a de comp endió que había
menosp eciado la in eligencia del ex
semina is a geo giano, y no había sido
capaz de alo a su genio pa a la in iga,
su des e güenza pa a men i y a ma
componendas.
Incapable o app ecia ing he Geo gian ex-
semina ian’s genius o in igue, his
shamelessness in lying and pu ing
oge he shady deals, Le Da ido ich
unde s ood oo la e ha he had
unde es ima ed his in elligence,
This me hod allows o mo e e icien manual e ision o he au oma ic alignmen ,
sa ing ime compa ed o a comp ehensi e manual e iew o each bisegmen . I s ikes a
balance be ween minimizing edious e iew wo k and main aining a high le el o
accu acy.
Figu e 5 ou lines he PaCo ES wo k low. Once he ex s ha e been aligned, hey a e
PoS- agged in each indi idual language be o e being ejoined and indexed. I should be
no ed he e ha due o echnical di icul ies, only he un agged e sions a e cu en ly
indexed despi e ha ing unde gone PoS- agging. The e o e, we do no add ess PoS-
agging in his pape .
Figu e 5: Wo k low o he c ea ion o he PaCo ES co e co po a
6 Sea ch and display ea u es
Once he ex s a e indexed, he esul ing da abase can be que ied wi h he help o he
SolR sea ch engine (Ve sion 7.5.0).
27
In his sec ion, a en ion is d awn o he ways he
use can que y he co pus and he kind o in o ma ion included in he sea ch esul s. In
bo h aspec s we ha e ied o mee he needs o a wide ange o use s, including expe s
and non-expe s.
6.1 Sea ch ea u es
The PaCo ES co po a can be que ied a h ee di e en le els: he i s le el
co esponds o he s anda d sea ch in e ace, he second o he ad anced sea ch
in e ace, and he hi d o an “expe le el”. Figu e 6 illus a es how he use can
al e na e be ween he s anda d and he ad anced sea ch modes in he case o he PaEnS
co pus.
27
h p://lucene.apache.o g/sol / and h ps://sol .apache.o g/docs/7_5_0/.
Raw ex
FL P ep ocessing
FL
Alignmen
FL<>ES
Pos- agging
FL
Pos- agging
ES
Indexa ion
pos agged
FL<>ES
Raw ex
ES P ep ocessing
ES
Figu e 6: Toggling be ween he s anda d and ad anced sea ch modes in he PaCo Es co po a
(PaEnS)
A he i s le el (s anda d sea ch in e ace), sea ches, by de aul , a e pe o med:
− based on lemma (i.e., he cha ac e chains a e in e p e ed as o ms o a lemma and
any segmen con aining o ms o ha lemma is added o he sea ch esul s). I
que ying he o m back, he use will be p esen ed wi h bo h segmen s con aining
back as a p eposi ion and segmen s con aining back as a noun. To sea ch o an
exac wo d, he sea ch e m mus be placed in quo a ion ma ks
− on he co e co pus, al hough he use can choose o include he supplemen s in he
que y by icking he espec i e boxes (see Figu e 6). This allows he use o
minimally selec he kind o language he o she wan s o sea ch o (usually mo e
gene al language in he co e co pus s. specialized language in he case o he
Eu opa l, TED, o Global Voices supplemen s, which con ain ins ances o
adminis a i e, popula science, and jou nalis ic language, espec i ely)
− wi hou di e en ia ing be ween o iginal and ansla ed ex s.
A he second le el (ad anced sea ch in e ace), he use can choose be ween he
ollowing op ions (see Figu e 7):
− que ying speci ic wo ks o g oups o wo ks by selec ing he speci ic ID numbe s,
au ho s, publica ion yea s, gen es, o a ie ies — US English, Mexican Spanish,
Aus ian Ge man, e c.
− in he case o he co e co pus, pe o ming he que ies on o iginal wo ks, ansla ed
wo ks o bo h. In he case o he Eu opa l supplemen , his can be done by selec ing
he co esponding language a ie y/ a ie ies in he ield “dialec al e sion”
− sea ching o a speci ic exp ession in one language which has (o does no ha e, see
below) a speci ic equi alen in he o he language, by placing each exp ession in he
co esponding language ield.
Figu e 7: Ad anced sea ch in e ace in he PaCo ES co po a (PaEnS)
The hi d o “expe ” le el does no ha e a speci ic sea ch in e ace. Expe le el
que ies can be pe o med bo h in he s anda d and in he ad anced sea ch in e ace. A
his le el, he use ( ypically a linguis ics, mode n language, o ansla ion s udies
schola ) can ake ull ad an age o he co po a by using he powe ul que y syn ax o
he unde lying que y ool SolR (Ve sion 7.5.0), and i needed, in combina ion wi h
egula exp essions (Regex). Some cla i ica ions a e pe inen he e:
− Expe sea ches can o cou se be combined wi h he possibili ies o ine- uned
que ies o e ed by he ad anced sea ch in e ace.
− The [SS] ope a o (meaning “SolR sea ch”) mus p ecede any que y chain a he
expe le el.
− The expe sea ch ope a o s a e also used in b owse s and bibliog aphic da abases,
so he expe use may al eady be acquain ed wi h hem. They include he wildca ds
* ( o mul iple cha ac e s) and ? ( o any single cha ac e ), he ilde ~ (ei he as he
uzzy sea ch ope a o and as pa o he dis ance ope a o ~1, ~2, e c.), as well as he
Boolean ope a o s NOT and OR ( he Boolean ope a o AND unde lies any sea ch
inpu consis ing o wo o mo e cha ac e -chains sepa a ed by a whi espace).
− On he Help page o each co pus a numbe o simple-wo d, mul i-wo d and
bilingual sample que ies a e displayed and explained in di e en ables o he use
o copy and pas e in o he sea ch ields. He o she can hen obse e he esul s and
build his o he own que ies by adap ing o enla ging he samples.
28
6.2 Display ea u es
Display ea u es a e also o ien ed o make in e ac ing wi h he PaCo ES co po a
amenable o di e en use g oups, as will be explained in his subsec ion.
On he one hand, he ollowing display ea u es a e o ien ed owa ds he gene al
use (an example o hese ea u es can be seen in Figu e 7, abo e):
− Ma ches a e shown in a side-by-side display ha is easie o p ocess isually han
he KWIC display ypical o monolingual co po a (Do al & Jiménez Juliá, 2019:
314-315).
− The le column co esponds o he o iginal ex and he igh column co esponds o
he ansla ed ex , be hey English o Spanish ( o example in he case o PaEnS);
− Ma ches appea ing in he o iginal ex s a e always shown i s .
− The que ied exp ession(s) appea s in bold, ega dless o whe he i appea s in he
o iginal ex s, in he ansla ed ex s, o in bo h.
− Ma ches a e p esen ed in g oups o 100; o a oid o e - ep esen a ion o a speci ic
wo k/au ho in he que y esul s, he ma ches a e andomized and shu led o each
que y. Tha means ha que ying he same exp ession wice o mo e will p esen he
use wi h di e en esul s each ime.
On he o he hand, he ollowing display ea u es migh be o in e es o speci ic,
mo e ad anced use g oups:
− Use s in ol ed in schola ly w i ing who include examples in hei ex s ob ained
om que ying he PaCo ES co po a will ind ha o each ma ch he bibliog aphical
e e ence o bo h he o iginal and he ansla ed pa o he bisegmen s (au ho ,
o iginal i le, ansla ed i le, place o publica ion o bo h o iginal and ansla ion,
and he pa and/o chap e o which he ma ch belongs) a e a ailable. This
in o ma ion can be accessed in a new window ia a sho link placed a e he
28
See, o example, he Help page a he PaGeS co pus:
h ps://www.co puspages.eu/co pus/help/help?lang=en& o ma =.
[11041] Cab e a In an e, Guille mo (2013). Mapa dibujado po un espía. Galaxia
Gu embe g.
English: Map d awn by a spy. A chipielago Books.
[11048] Padu a, Leona do (2009/ 2013). El homb e que amaba a los pe os. Tusque s.
English: The man who lo ed Dogs. Bi e Lemon P ess.
Add ess o co espondence:
Ma ía Te esa SÁNCHEZ NIETO
Uni e sidad de Valladolid
Facul ad de T aducción e In e p e ación
Campus Uni e si a io Duques de So ia, s/n
E- 42003 SORIA (Spain)
ma ia e esa.sanchez[email p o ec ed]
O cidID: 0000-0002-0378-3720
Co-au ho :
I ene DOVAL REIXA
Uni e sidade de San iago de Compos ela
Facul ad de Filología
A da. de Cas elao, s/n,
E-15782 SANTIAGO DE COMPOSTELA (Spain)
[email p o ec ed]
O cidID: 0000-0002-9050-3183
