Improving a Catalan-Spanish Statistical Translation System using Morphosyntactic Knowledge

Proceedings of the 13th Annual Conference of the EAMT, pages 52–57, Barcelona, May 2009 Improving a Catalan-Spanish Statistical Translation System using Morphosyntactic Knowledge Mireia Farrús, Marta R. Costa-jussà, Marc Poch, Adolfo Hernández, and José B. Mariño TALP Research Centre, Department of Signal Theory and Communications Universitat Politècnica de Catalunya C/ Jordi Girona 1-3, 08034 Barcelona, Spain {mfarrus,mruiz,mpoch,adolfohh,canton}@gps.tsc.upc.edu Abstract In this paper, a human evaluation of a Catalan-Spanish Ngram-based statistical machine translation system is used to de- velop specific techniques based on the use of grammatical categories, lexical cate- gorisation and text processing, for the en- hancement of the final translation. The system is successfully improved when test- ing with ad hoc and general corpora, as it is shown in the final automatic evaluation. 1 Introduction Statistical Machine Translation (SMT) nowadays has become one of the most popular Machine Translation paradigms. The SMT approach allows to build a translator with open-source tools as long as a parallel corpus is available. If the languages involved in the translation belong to the same lin- guistic family, the translation quality can be sur- prisingly nice. Furthermore, one of the most at- tractive reasons to build an statistical system in- stead of an standard rule-based system is the little human effort required. Theoretically, when using SMT, no linguistic knowledge is required. In practice, once the sys- tem is built and specially, if the translation qual- ity is high, then the linguistic knowledge becomes necessary to make further improvements (Niessen and Ney, 2000; Popovi´c and Ney, 2004; Popovi´c et al., 2006). In fact, the main question that arose at the beginning of this work was: which are the steps to follow when the intention is to improve a high quality statistical translation? Let’s consider a high quality statistical trans- lation defined as the system which has a BLEU c⃝ 2009 European Association for Machine Translation. around 75% with a single reference in an in- domain test. This is a relatively unusual situation as most of the statistical translation systems have much lower performance. This study is devoted to develop this stage in the Catalan-Spanish pair in both directions. The study starts from a high quality Ngram- based statistical translation baseline system, trained with the aligned Spanish-Catalan parallel corpus taken from El Periódico newspaper, which contains 1.7 million sentences. A human error analysis of the translation is then performed and used to further improve the translation by introduc- ing statistical techniques and linguistic rules. This paper is organised as follows. Section 2 de- scribes the Ngram-based statistical translation sys- tem used as baseline system. Section 3 reports the human error analysis and evaluation of the base- line system, whose solutions based on statistical techniques, linguistic rules and text processing are explained in section 4. In section 5, an automatic evaluation of the new system is performed and dis- cussed. Finally, Section 6 sums up the conclusions. 2 Ngram-based statistical translation system An Ngram-based SMT system regards translation as a stochastic process. In recent systems, such an approach …

Content truncated to 500 words