Rule-Based Augmentation of Training Data in Breton-French Statistical Machine Translation

Proceedings of the 13th Annual Conference of the EAMT, pages 213–217, Barcelona, May 2009 Rule-Based Augmentation of Training Data in Breton–French Statistical Machine Translation Francis M. Tyers Dept. de Llenguatges i Sist. Inform`atics, Universitat d’Alacant E-03071 Alacant (Spain) Prompsit Language Engineering Av. St. Francesc d’Ass´ıs, 74, 1r-L E-03195 l’Altet (Spain) ftyers@prompsit.com Abstract This article describes an initial statistical machine translation system between Bre- ton, a Celtic language spoken in France, and French. It also describes a method for leveraging existing resources from an incomplete rule-based machine translation system to improve the coverage and trans- lation quality of the statistical system by generating expanded bilingual vocabulary lists. Results are presented which show that the use of this method improves the re- sults of the system with respect to both the baseline, and the baseline with a lemma- to-lemma bilingual lexicon. 1 Introduction Breton is a Celtic language of the Brythonic branch largely spoken in Brittany in the north- west of France. Historically it was spoken only in the northern part of Brittany, Breizh-Izel (Lower Brittany). This contrasts with Breizh-Uhel (Higher Brittany) which is traditionally Romance- speaking. Although some sources put the number of native speakers at between 500,000 and 600,000 (Gor- don, 2005), a more up-to-date estimate can be found from the organisation Ya d’ar brezhoneg which gives a number of 201,083 as of the 11th November 2008, and states that the number is de- creasing at a rate of at least one per hour.1 Breton is classed as a language in serious danger of ex- tinction by the UNESCO Red Book on Endangered Languages (Salminen, 1999), a situation exacer- c⃝ 2009 European Association for Machine Translation. 1http://www.yadarbrezhoneg.com/ ?article245; Accessed: 11th November, 2008 bated by the laissez-faire policies of the French state. Like other Celtic languages, Breton exhibits the phenomenon of initial consonant mutation. This occurs when the initial consonant of a word changes based on morpho-syntactic context. For example in the word tad “father”, the initial con- sonant mutates to a ‘z’ (aspirant mutation) when the word follows the possessive ma “my”, so tad is “father”, while “my father” is ma zad. As for many less-resourced language pairs, while there is little aligned bilingual text, bilin- gual lexicons are more readily available. One so- lution would be to use these bilingual lexicons within a rule-based system that makes use of the features found in the bilingual lexicon: part-of- speech, gender, number etc. to try to compensate for the lack of data with some level of generalisa- tion. Even if little parallel data is available, it is still worthwhile to compare any attempt at a more linguistically motivated system with a greater gen- eralising power with a straight-forward, state-of- the-art non-linguistic approach, such as phrase- based statistical machine translation (SMT). 2 Resources 2.1 Parallel corpora For any language pair, parallel corpora are the scarcest of all resources. In the case of a lan- guage with a small population of speakers and no official recognition, the hurdle is even greater. In contrast …

Content truncated to 500 words