Noisy SMS Machine Translation in Low-Density Languages

Proceedings of the 6th Workshop on Statistical Machine Translation, pages 344–350, Edinburgh, Scotland, UK, July 30–31, 2011. c⃝2011 Association for Computational Linguistics Noisy SMS Machine Translation in Low-Density Languages Vladimir Eidelman†, Kristy Hollingshead†, and Philip Resnik†‡ †UMIACS Laboratory for Computational Linguistics and Information Processing ‡Department of Linguistics University of Maryland, College Park {vlad,hollingk,resnik}@umiacs.umd.edu Abstract This paper presents the system we developed for the 2011 WMT Haitian Creole–English SMS featured translation task. Applying stan- dard statistical machine translation methods to noisy real-world SMS data in a low-density language setting such as Haitian Creole poses a unique set of challenges, which we attempt to address in this work. Along with techniques to better exploit the limited available train- ing data, we explore the benefits of several methods for alleviating the additional noise inherent in the SMS and transforming it to better suite the assumptions of our hierarchi- cal phrase-based model system. We show that these methods lead to significant improve- ments in BLEU score over the baseline. 1 Introduction For the featured translation task of the Sixth Work- shop on Statistical Machine Translation, we devel- oped a system for translating Haitian Creole Emer- gency SMS messages. Given the nature of the task, translating text messages that were sent during the January 2010 earthquake in Haiti to an emergency response service called Mission 4636, we were not only faced with the problem of dealing with a low- density language, but additionally, with noisy, real- world data in a domain which has thus far received relatively little attention in statistical machine trans- lation. We were especially interested in this task be- cause of the unique set of challenges that it poses for existing translation systems. We focused our re- search effort on techniques to better utilize the lim- ited available training resources, as well as ways in which we could automatically alleviate and trans- form the noisy data to our advantage through the use of automatic punctuation prediction, finite-state raw-to-clean transduction, and grammar extraction. All these techniques contributed to improving trans- lation quality as measured by BLEU score over our baseline system. The rest of this paper is structured as follows. First, we provide a brief overview of our baseline system in Section 2, followed by an examination of issues posed by this task and the steps we have taken to address them in Section 3, and finally we con- clude with experimental results and additional anal- ysis. 2 System Overview Our baseline system is based on a hierarchical phrase-based translation model, which can formally be described as a synchronous context-free gram- mar (SCFG) (Chiang, 2007). Our system is imple- mented in cdec, an open source framework for align- ing, training, and decoding with a number of differ- ent translation models, including SCFGs. (Dyer et al., 2010). 1 SCFG grammars contain pairs of CFG rules with aligned nonterminals, where by introduc- ing these nonterminals into the grammar, such a sys- tem is able to utilize both word and phrase level re- ordering to capture …

Content truncated to 500 words