BP2EP - Adaptation of Brazilian Portuguese texts to European Portuguese

BP2EP – Adaptation of Brazilian Portuguese texts to European Portuguese Lu´ıs Marujo 1,2,3, Nuno Grazina 1, Tiago Lu´ıs 1,2, Wang Ling 1,2,3, Lu´ısa Coheur 1,2, Isabel Trancoso 1,2 1 Spoken Language Laboratory / INESC-ID Lisboa, Portugal 2 Instituto Superior T´ecnico, Lisboa, Portugal 3 Language Technologies Institute / Carnegie Mellon University, Pittsburgh, USA {luis.marujo,ngraz,tiago.luis,wang.ling,luisa.coheur,imt}@l2f.inesc-id.pt Abstract This paper describes a method to effi- ciently leverage Brazilian Portuguese re- sources as European Portuguese resources. Brazilian Portuguese and European Por- tuguese are two Portuguese varieties very close and usually mutually intelligible, but with several known differences, which are studied in this work. Based on this study, we derived a rule based system to translate Brazilian Portuguese resources. Some resources were enriched with mul- tiword units retrieved semi-automatically from phrase tables created using statisti- cal machine translation tools. Our experi- ments suggest that applying our translation step improves the translation quality be- tween English and Portuguese, relatively to the same process using the same re- sources. 1 Introduction Modern statistical machine translation (SMT) depends crucially on large parallel corpora and on the amount and specialization of the training data for a given domain. Depending on the domain, such resources may sometimes be available in only one of the language varieties. For instance, class- room lecture and talk transcriptions such as the ones that one can find in the MIT OpenCourse- Ware (OCW) website1, and the TED talks website2 (838 BP vs 308 EP) are examples of parallel cor- pora where Brazilian Portuguese (BP) translations can be found much more frequently than European c⃝ 2011 European Association for Machine Translation. 1http://ocw.mit.edu/OcwWeb 2http://www.ted.com/talk Portuguese (EP) translations. This discrepancy has origin in the Brazilian population size that is near 20 times larger than the Portuguese population. This paper describes the progressive develop- ment of a tool that transforms BP texts into EP, in order to increase the amount of EP parallel corpora available. This paper is organized as follows: Section 2 de- scribes some related work; Section 3 presents the main differences between EP and BP; the descrip- tion of the BP2EP system is the focus of Section 4. The following section describes an algorithm for extracting Multiword Lexical Contrastive pairs from SMT Phrase-tables; Section 6 presents the re- sults, and Section 7 concludes and suggests future work. 2 Related Work Few papers are available on the topic of improv- ing Machine Translation (MT) quality by explor- ing similarities in varieties, dialects and closely re- lated languages. Altintas (2002) states that developing an MT system between similar languages is much easier than the traditional approaches, and that by putting aside issues like word reordering and most of the semantics, which are probably very similar, it is possible to focus on more important features like grammar and the translation itself. This also al- lows the creation of domains of closely related lan- guages which may be interchangeable and that, in this particular case, would allow, for instance, the development of MT systems between English and a set of Turkic …

Content truncated to 500 words