The BM-I2R Haitian-Créole-to-English translation system description for the WMT 2011 evaluation campaign

Proceedings of the 6th Workshop on Statistical Machine Translation, pages 452–456, Edinburgh, Scotland, UK, July 30–31, 2011. c⃝2011 Association for Computational Linguistics The BM-I2R Haitian-Cr´eole-to-English translation system description for the WMT 2011 evaluation campaign Marta R. Costa-juss`a Barcelona Media Innovation Center Av Diagonal, 177, 9th floor 08018 Barcelona marta.ruiz@barcelonamedia.org Rafael E. Banchs Institute for Infocomm Research 1 Fusionopolis Way 21-01 Singapore 138632 rembanchs@i2r.a-star.edu.sg Abstract This work describes the Haitian-Cr´eole to En- glish statistical machine translation system built by Barcelona Media Innovation Center (BM) and Institute for Infocomm Research (I2R) for the 6th Workshop on Statistical Ma- chine Translation (WMT 2011). Our sys- tem carefully processes the available data and uses it in a standard phrase-based system en- hanced with a source context semantic feature that helps conducting a better lexical selection and a feature orthogonalization procedure that helps making MERT optimization more reli- able and stable. Our system was ranked first (among a total of 9 participant systems) by the conducted human evaluation. 1 Introduction During years there has been a big effort to produce natural language processing tools that try to under- stand well written sentences, but the question is how well do these tools work to analyze the contents of SMS. For example, not even syntactic tools like stemming can bring to common stems words that have been shortened (like Xmas or Christmas). This paper describes our participation on the 6th Workshop on Statistical Machine Translation (WMT 2011). The featured task from the workshop was to translate Haitian-Cr´eole SMS messages into En- glish. According to the WMT 2011 organizers, these text messages (SMS) were sent by people in Haiti in the aftermath of the January 2010 earthquake. Our objective in this featured task is to translate from Haitian-Cr´eole into English either using raw or clean data. We propose to build an SMT system which could be used for both raw and clean data. Our base- line system is an standard phrase-based SMT sys- tem built with Moses (Koehn et al., 2007). Starting from this system we propose to introduce a semantic feature function based on latent semantic indexing (Landauer et al., 1998). Additionally, as a total dif- ferent approximation, we propose to orthogonalize the standard feature functions of the phrase-based table using the Gram-Schmidt methodology (Greub, 1975). Then, we experimentally combine both en- hancements. The only difference among the raw and clean SMT system were the training sentences. In order to translate the clean data, we propose to normalize the corpus of short messages given very scarce re- sources. We only count with a small set of parallel corpus at the level of sentence of chat and standard language. A nice normalization methodology can allow to make the task of communication easier. We propose a statistical normalization technique using the scarce resources we have based on a combina- tion of statistical machine translation techniques. The rest of this paper is organized as follows. Sec- tion 2 briefly describes the phrase-based SMT sys- tem which is used as a …

Content truncated to 500 words