apertium-cy - a collaboratively-developed free RBMT system for Welsh to English

apertium-cy (http://www.cymraeg.org.uk) is a rule-based “gisting” machine translation system forWelsh to English, with both engine and data released under the GPL.We summarise the development of apertium-cy, evaluate its output, and discuss the advantages of a collaborative development model combined with rule-based MT for marginalised languages. 1. e Apertium platform apertium-cy is a “gisting” machine translation system for Welsh to English, based on the Apertium machine translation platform.1 e platform was originally aimed at the Romance languages of the Iberian peninsula, but is now being adapted for other languages (such as Basque, and languages from the Celtic group –Welsh, Irish, Breton), withmuch of the work on new languages being pursued by volunteers, following the increasingly common collaborative development model used for free2 and open-source soware. Apertium is licensed under the Free Soware Foundation’s GNUGeneral Public License,3 and all the soware and data for the 17 supported language pairs (and the other pairs in development) is available for download from the project website. Apertium follows a shallow-transfer approach, and is very fast. Finite-state transducers (Garrido-Alenda and Forcada, 2002, Roche and Schabes, 1997) processing up to 40,000 words per second are used for lexical processing, first-order hiddenMarkov models (HMM) are used for part-of-speech tagging, and multi-stage finite-state based chunking for structural transfer. 1http://www.apertium.org 2We follow the definition of “free” used by the Free Soware Foundation http://www.fsf.org. 3http://www.fsf.org/licensing/licenses/gpl.html, accessed 12/12/2008. © 2009 PBML. All rights reserved. Please cite this article as: Francis Tyers, Kevin Donnelly, apertium-cy – a collaboratively-developed free RBMT system for Welsh to English. The Prague Bulletin of Mathematical Linguistics No. 91, 2009, 57–66. PBML 91 JANUARY 2009 e soware behind the platform is implemented as a standard UNIX pipeline, with each stage in the translation having a separate C++ program. Communication between each stage uses piped text streams. XML-based formats are used to encode the linguistic data, which are then compiled into the high-speed formats used by the engine. Further details are given in (Armentano-Oller et al., 2006), and on the project website.