Automatic translation and Google's Summer of Code

Automatic translation from Welsh gets a boost from France!

High-quality Welsh-English machine translation will come a step closer when a new initiative gets underway this month.

The multinational Apertium team, which released their Welsh-English translator (http://www.cymraeg.org.uk [1]) in August 2008, has been accepted into the fifth Google Summer of CodeTM [2], and one of the projects to be funded will be an improvement to that translator.

Apertium (http://www.apertium.org) is a Free Software [3] machine translation platform. It was first developed to handle translation between related languages in Spain, but over the last few years it has been extended to deal with other languages. To date, translators for 17 language pairs have been released, covering languages spoken by 1.1bn people, from English (est. 500m speakers) to Aranese (est. 4,000 speakers). A similar number of other language pairs are in development - these include Indian languages like Hindi and Bengali, and Scandinavian languages like Norwegian and Sami.

Google Summer of Code offers student developers stipends to write code for open-source projects, advised by mentors already working on the projects, and has helped create millions of lines of code for dozens of projects. This was the first year that Apertium applied for the program, and 9 Apertium projects are being supported.

The Apertium Welsh-English translator works by applying grammatical rules to a Welsh sentence to turn it into an English sentence. An alternative approach (adopted by software like Moses [4]) is to use a large body of text to work out what the likely translation of a given phrase is.

The Summer of Code student, Gabriel Synnaeve from Grenoble, France [5], will be working on combining these two approaches, using techniques developed at Carnegie-Mellon University in the USA [6]. The aim is to improve the quality of the translation - in effect, the Apertium and Moses translations will be compared, and the best bits of each will be used in the final translation.

For instance, take the Welsh sentence: "Mae Heddlu'r De yn ymchwilio i farwolaeth dyn 41 oed o Abertawe." (South Wales Police are investigating the death of a 41-year old man from Swansea.)

Apertium currently produces: "South Wales Police is investigating death man 41 years old from Swansea."

Moses currently produces: "the south wales police investigation into the death of a man 41 years of age of abertawe."

The aim is to combine the best chunks from each program, so that we get something like: *[South Wales Police] *[is investigating] +[the death of a man] *[41 years old] *[from Swansea] Here, the chunks marked * come from Apertium, and the one marked + from Moses, and combining both improves the quality of the translation.

This is cutting-edge stuff, and has rarely been tried before. Prof Harold Somers, in a 2004 report for the Welsh Language Board [7], suggested that a medium-term goal for machine translation in Welsh would be “to integrate ... different [machine translation] engines into a single system”. Nothing has been done on that to date, and Gabriel's work will be the first attempt to bring this vision of "multi-engine machine translation" for Welsh closer to reality.

Francis Tyers [8], who will be mentoring Gabriel, said, "I was quite surprised that we didn't get any Welsh students applying, but this is a fantastic opportunity to improve Welsh language technology. I have no doubt we'll see some real gains in the translation quality."

Gabriel has already started work. "At the minute I'm fine-tuning the Moses Welsh-English translator to make it as efficient as possible. The Apertium community is very friendly, and I wanted to participate in a big open source project, so I'm glad I went for it."

Kevin Donnelly [9], who co-developed the Apertium Welsh-English translator with Francis, noted that this was a big step forward for Welsh. "It is wonderful that so many talented people are working on Apertium, and that they are giving Welsh such a high priority. What we need now is for bodies promoting Welsh here in Wales to step up to the plate and give whatever enouragement and other support they can."

Notes

[1] http://ufal.mff.cuni.cz/pbml-91-100.html. Francis Tyers and Kevin Donnelly (2009): "apertium-cy - a collaboratively-developed free RBMT system for Welsh to English", Prague Bulletin of Mathematical Linguistics, 91.

[2] http://code.google.com/soc

[3] http://www.fsf.org/about/what-is-free-software. The Free Software Foundation's definition of "Free Software" is software that the user is free to use, copy, change, and distribute.

[4] http://www.statmt.org/moses. Moses is an open-source statistical machine translation system.

[5] Gabriel Synnaeve is a student at the École Nationale Supérieure d'Informatique et de Mathématiques (http://ensimag.grenoble-inp.fr), a leading informatics and mathematics centre. He will graduate in September 2009 and will then begin work on a doctorate on Bayesian machine learning.

[6] Alon Lavie (http://www.cs.cmu.edu/~alavie) is leading this work. See also: http://www.cs.cmu.edu/~alavie/papers/EAMT-2005-MEMT.pdf. S. Jayaraman and A. Lavie (2005): "Multi-Engine Machine Translation Guided by Explicit Word Matching", Proceedings of EAMT-2005.

[7] http://www.byig-wlb.org.uk/english/publications/publications/2302.doc. Harold Somers (2004): "Machine translation and Welsh: the way forward.", Report for the WLB.

[8] Francis Tyers studied computer science at Aberystwyth, and is now a language engineer for Prompsit Language Engineering, S.L. and a PhD student at the Universitat d'Alacant. He is a key Apertium developer, with a special interest in extending it to handle the Celtic languages.

[9] Kevin Donnelly has been working on Free Software in Welsh since 2003, and developed the online Welsh dictionary Eurfa (http://www.eurfa.org.uk).