Apertium: a unique free/open-source MT system for related languages [but not only]
1. #LocWorld34
Apertium: a Unique
Free/Open-Source MT System
for Related Languages
[but not only]
Gema Ramírez Sánchez1
Mikel L. Forcada1,2
1
Prompsit Language Engineering, Elx, Spain
1,2
Universitat d’Alacant, Alacant, Spain
2. #LocWorld34
Outline
● Apertium components
● Ready-to-use Apertium products
● Machine translation — but not only!
● Licensing — free/open-source
● The Apertium community
● Research and business with Apertium
● Languages and language pairs
● Success cases
● Funding
4. #LocWorld34
Apertium components: the
engine /1
● A fast, free/open-source, modular,
shallow-transfer, language-independent
machine translation engine with:
○ text format management,
○ translation memory querying,
○ finite-state lexical processing,
○ statistical and constraint-based lexical
disambiguation, and
○ shallow structural transfer based on
finite-state pattern matching
5. #LocWorld34
Apertium components: the
engine /2
● Most of the engine was developed inside the
Apertium project but some external
technologies are used:
○ Helsinki Finite-State toolkit (for some
morphologically-rich languages),
○ VISL CG-3 (constraint grammars for
rule-based lexical disambiguation).
6. #LocWorld34
Apertium components: the data
● Free/open-source language data in
well-specified XML formats for a variety of
languages and language pairs.
7. #LocWorld34
Apertium components: the data.
A typical language pair
Language pair organization
2 monolingual packages (A, B)
▪ 1 monolingual dictionary
(monodix)
▪ 1 tagset + probabilities
▪ 1 plain/tagged corpus
▪ 1 postgeneration “dictionary”
1 bilingual package (A–B)
▪ 1 bilingual dictionary
(bidix)
▪ 2 sets of structural transfer
(grammar) rules (levels 1–3)
Format: typically XML-based (sometimes text-based) files
Sizes:
Monodixes: 10k–90k lemmata; 100k–23M surf. forms, 85–97% cover.
Bidixes: 8k--–90k bilingual lema correspondences
Rules: 100 (one level) – 300 (3 level) per translation direction
8. #LocWorld34
Apertium components: the
tools
● Free/open-source tools:
○ compilers to turn linguistic data into a fast
and compact form used by the engine
and
○ software to learn disambiguation or
translation rules from corpora.
9. #LocWorld34
Ready-to-use Apertium
products
● A stand-alone Java application for the
desktop: apertium-caffeine.
● An Android version for handhelds.
● A stand-alone version (Apertium Simpleton)
for Windows and MacOS.
● Plug-ins and support for CAT platforms:
OmegaT, MateCat, MemoQ, Trados Studio.
● Available as a PPA repository for GNU/Linux
users.
12. #LocWorld34
No need to install: web access
www.apertium.org
● Text box: short plain texts
● Document translation:
○ plain text
○ HTML, XML (.xliff)
○ OpenDocument (.odt, .odp, .ods)
○ Office “-x” formats: .docx, .xlsx, .pptx
○ LaTeX
● A nice feature: with/without marks for
unknown words
13. #LocWorld34
No need to install: web/API
access
● Other portals with all Apertium languages:
○ Prompsit’s portal: + TMX +
navigate&translate
○ iTranslate4.eu portal: multiengine
● Other portals with some Apertium languages:
○ UOC, UPV, UA (+ TMX + terminology support
+ more formats)
○ GiellaTekno portal
○ etc.
● Also API access and connectors to translation
tools are marketed
15. #LocWorld34
Machine translation — but
not only! /2
Monodix
Tagset+prob
Rules
Monodix
Bidix
t
o
o
l
s
t
o
o
l
s
Post-dix
Morphological
analyser
PoS tagger
Lexical transfer
Full MT
Morphological
generator
Structural transfer
Post-generator
16. #LocWorld34
Machine translation — but
not only! /3
● Apertium is a rule-based machine translation
system but the pipeline contains many
monolingual modules that can be used for other
human-language technology tasks (such as
anonymization or factored output)
● Most modules are based on finite-state
technology; HMMs are used for part-of-speech
tagging and an interpreted language is used to
write structural transfer rules.
17. #LocWorld34
Licensing: free/open-source /1
Apertium language data and code are both
licensed under the GNU General Public License:
● a free/open-source license allowing free
distribution of unmodified and modified
versions
● a copylefted license: it avoids private
appropriation and encourages giving
improvements back to the project (it creates
a software commons).
18. #LocWorld34
Licensing: free/open-source /2
● The free/open-source model creates a
community which effectively connects
researchers, developers, vendors, and users
in a continuum.
19. #LocWorld34
The Apertium community
● Very active group of hundreds of developers
● Contributions to Apertium at Sourceforge
● Wiki documentation (wiki.apertium.org)
● Easy entry: Apertium linguistic modelling is
simple, no need to program.
● IRC channel #apertium in freenode.net
● Mailing lists: apertium-stuff@lists.sf.net and
other lists
23. #LocWorld34
The Apertium community:
activities
● President and project management
committee election according to bylaws
● Support: mail, chat, online meetings
● Maintenance: pairs, web, mobile app
● Manuals & documentation: wiki, manuals,
how-to’s, training materials
● Organization of Google Summer of Code and
Google Code-In activity
● Outreach activities: conferences, workshops
● Language-related groups
24. #LocWorld34
Research and business with
Apertium
Apertium is already an active research and
business platform:
● Research: 40+ publications, 2 PhD thesis, 4
master's theses.
● Business: companies (Prompsit, Eleka,
Imaxin Software, etc.) offering services to
customers such as Autodesk, Adobe, the
Government of Catalonia, 2 daily newspapers
in Spain, freelancers and LSPs
25. #LocWorld34
Languages and language pairs
/1
● Language data is encoded mostly in XML,
but some language pairs contain data
encoded in other text-based formats.
● Stable language pairs (bilingual data) are
currently more than 40.
28. #LocWorld34
Languages and language pairs
/4
Year Milestone Language pairs
2004 The Spanish Ministry of Industry funds a consortium
to build FOSS MT for the languages of Spain ----------------------------
2005 Apertium RBMT plaftorm is launched providing
engine, tools and data under free licenses
3 pairs: es–ca, es–gl
and es–pt
2005-2009 Language pair-driven innovation, still very
European-focused language pairs
+19: fr, en, eo, ro, eu,
oc, cy, nn, nb, sv, da, is,
mk, bg, ast, br
2010 Five years on! 22 pairs!!!
2011-2015 Consolidated community, support for non-European
languages, new tools and reorganisation of data
+19: af, nl, hr, sr, mt, sl,
arg, sme, urd, hin, kaz,
tat, id, ms, ar
2017 Twelve years on! 43 pairs!!!
29. #LocWorld34
Apertium loves small
languages
● Breton→French
● Aragonese↔Spanish/Catalan
● Occitan↔Catalan/Spanish
● Italian→Sardinian
● North Sámi↔Norwegian
● Icelandic↔Swedish
● Spanish→Spanish Sign Language
30. #LocWorld34
Language pairs with approx.
95% text coverage
Language Lemmata Inflection models Surface forms
HBS 97,445 1,429 23,348,650
English 60,543 312 108,119
Spanish 46,003 442 4,737,777
Catalan 41,116 559 7,088,585
Galician 29,818 333 14,247,591
Asturian 46,550 443 18,541,752
Occitan 21,602 527 6,084,575
Aragonese 26,068 544 12,870,976
Portuguese 14,436 316 10,514,672
31. #LocWorld34
Apertium language-pair
life cycles
● For new pairs:
○ resource compilation
○ basic system creation (85% coverage, most
frequent structural phenomena)
○ evaluation
○ typically takes 3–6 months
● For existing pairs:
○ testing, enhancement, evaluation
○ typically takes 1–3 months
32. #LocWorld34
A related-languages pair
performance: apertium-es-pt
From Masselot et al., 2010 (Using the Apertium
Spanish–Brazilian Portuguese MT system for
localization):
● Post-editing effort (word error rate): 20%
● Post-editing speed: average 4,500 words/day
Updated 2017 (also for software localisation):
● Post-editing effort (word error rate): 14%
● Post-editing speed: average 6,500 words/day
34. #LocWorld34
Related language-pair
post-editing experience /2
Original Spanish Apertium output Portuguese final
Cree documentación
y dibujos 2D con un
completo conjunto
de herramientas de
dibujo, edición y
anotación.
Crê documentação e
desenhos 2D com
um completo
conjunto de
ferramentas de
desenho, edição e
anotação.
Produza desenhos e
documentação 2D
com um conjunto
abrangente de
ferramentas de
desenho, edição e
anotação.
Apertium output for closely-related languages is:
● Easy and fast to post-edit
● Rather mechanical, but reliable
● Predictable
35. #LocWorld34
Nearby LocWorld Barcelona...
● Apertium makes two daily newspaper
bilingual: Levante (Catalan) and La Voz de
Galicia (Galician).
● Universities in the Catalan speaking area
use Apertium to help in the generation of
courseware and academic information;
● Apertium is used in PLATA, the Spanish
government platform for webpage
translation.
Some success cases /1
36. #LocWorld34
Also by-products:
● Same-language machine translation for
local flavours/flavors: AltLang.net
○ available for English, Spanish, French and
Portuguese varieties.
○ performs spelling, lexical, grammar and
style changes.
Some success cases /2
Based on Apertium
38. #LocWorld34
Wikimedia Content Translation
into Norwegian Nynorsk
Co-funded project on MT
for Scandinavian
languages including
community outreach starts
Most of the translations are
from Norwegian Bokmål.
85% are done using Apertium.
40. #LocWorld34
● Translators Without Borders develop
crisis-specific, portable machine translation
from English to Kurdish languages (Kurmanji,
Sorani) on Apertium.
● Apertium and language experts help promote a
unified standard for Occitan by defining and
selecting it for Spanish→Occitan and
Catalan→Occitan MT
Other success cases involving
interaction with other
communities
41. #LocWorld34
Funding /1
● The Ministry of Industry, Tourism and
Commerce of Spain (also, the Ministries of
Education and Science and of Science and
Technology of Spain)
● The Secretariat for Technology and the
Information Society of the Government of
Catalonia
● The European Commission (DGT training and
Abu-Matran project)
● The Ministry of Foreign Affairs of Romania
42. #LocWorld34
Funding /2
● Universitat d'Alacant and Universitat Oberta
de Catalunya
● Ofis Publik ar Brezhoneg (Breton Language
Board)
● Ministry of Education and Science of the
Republic of Kazakhstan
● Google Summer of Code scholarships
(2009–2014, 2016, 2017) and Google Code-In
donations (2010–2016).
● And many other private companies
43. #LocWorld34
● If you want to build, integrate, or customize
fast, reliable, predictable machine translation for
your application.
● If you’d rather understand application-oriented
dictionaries and rules rather than deal with the
“magic” of embeddings, decoders, phrase tables,
convolutions, or probabilities.
● If there’s no way you can amass and curate
millions of translated words to train a system
for your language or application.
Then come and talk to us
(we are at booth 121).
You can be part of it!