Speeding up the Implementation Process of a Shallow Transfer Machine Translation System

Speeding up the Implementation Process of a Shallow Transfer Machine Translation System Jernej Viˇciˇc University of Primorska, PINT Glagoljaška 8, Koper Slovenia jernej.vicic@upr.si Petr Homola Institute of Formal and Applied Linguistics Malostranské námˇestí 25, Praha, Czech Republic homola@ufal.mff.cuni.cz Abstract The article presents an attempt to auto- mate all data creation processes of a rule- based shallow-transfer machine translation system. The presented methods were tested on two fully functional translation systems Slovenian-Serbian and Slovenian- Macedonian. An extensive range of evalu- ation tests was performed to assess the ap- plicability of the methods. 1 Introduction and problem statement Several methods that automate some parts of the shallow transfer Rule Based Machine Translation (RBMT) system construction have been presented and are even used as part of the construction toolkits like Apertium (Corbi-Bellot et al., 2005), which is a widely used architecture for creating machine translation systems between related lan- guages. (Hajiˇc et al., 2000), (Hajiˇc et al., 2003) and (Corbi-Bellot et al., 2005) suggest using an ar- chitecture similar to the one presented in Figure 1. All methods and materials discussed in this paper were tested on a fully functional machine transla- tion system based on Apertium. The construction of a machine translation system for a new language pair falls into roughly two cat- egories: • A job of manual dictionary and rule con- struction for Rule-Based Machine Translation (RBMT) system construction. • Automatic machine translation system con- struction in case of corpus-based machine construction systems such as Statistical Ma- chine Translation (SMT) (Brown et al., 1993; c⃝ 2010 European Association for Machine Translation. Och and Ney, 2003) or Example-Based Ma- chine Translation (EBMT) (Nagao, 1984; Hutchins, 2005). The SMT seems like a perfect choice as some of the best performing machine translation sys- tems are based on the SMT technologies (NIST, 2006), but it has a few drawbacks that cannot be ignored; the SMT systems, to be efficient, re- quire huge amount of parallel text (Och, 2006) that is available only for a few of the widely used languages like English, Spanish, French, Arabic, etc. The morphologically rich and highly inflec- tive languages like the pairs presented in this paper (Slovenian paired with Serbian and Macedonian) present an even bigger problem. One of the most appealing reasons for using an RBMT machine translation system is the ability for the experts of the field to further refine the re- sults of the automatically produced data. An attempt to fully automatise the construction of all the data for a fully functional shallow transfer RBMT system has been presented in (Viˇciˇc, 2008). Parts of the creation process have been addressed by several authors such as automated monolingual dictionary extraction (Forsberg et al., 2006); sup- port for agglutinative languages (Beesley and Kart- tunen, 2000); Part Of Speech (POS) tagger train- ing (Sánchez-Martínez et al., 2008; Halácsy et al., 2007; Brants, 2000); automatic induction of shallow-transfer rules (Sanchez-Martinez and For- cada, 2009); automatic extraction of bilingual dic- tionaries (Caseli et al., 2008). Some of these tech- nologies …

Content truncated to 500 words