CMU Haitian Creole-English Translation System for WMT 2011

Proceedings of the 6th Workshop on Statistical Machine Translation, pages 386–392, Edinburgh, Scotland, UK, July 30–31, 2011. c⃝2011 Association for Computational Linguistics CMU Haitian Creole-English Translation System for WMT 2011 Sanjika Hewavitharana, Nguyen Bach, Qin Gao, Vamshi Ambati, Stephan Vogel Language Technologies Institute Carnegie Mellon University Pittsburgh, PA 15213, USA {sanjika,nbach,qing,vamshi,vogel+}@cs.cmu.edu Abstract This paper describes the statistical machine translation system submitted to the WMT11 Featured Translation Task, which involves translating Haitian Creole SMS messages into English. In our experiments we try to ad- dress the issue of noise in the training data, as well as the lack of parallel training data. Spelling normalization is applied to reduce out-of-vocabulary words in the corpus. Us- ing Semantic Role Labeling rules we expand the available training corpus. Additionally we investigate extracting parallel sentences from comparable data to enhance the available par- allel data. 1 Introduction In this paper we describe the CMU-SMT Haitian Creole-English translation system that was built as part of the Featured Translation Task of the WMT11. The task involved translating text (SMS) messages that were collected during the humanitarian opera- tions in the aftermath of the earthquake in Haiti in 2010. Due to the circumstances of this situation, the SMS messages were often noisy, and contained in- complete information. Additionally they sometimes contained text from other languages (e.g. French). As is typical in SMS messages, abbreviated text (as well as misspelled words) were present. Further, since the Haitian Creole orthography is not fully standardized (Allen, 1998), the text inherently con- tained several different spelling variants. These messages were translated into English by a group of volunteers during the disaster response. The background and the details of this crowdsourc- ing translation effort is discussed in Munro (2010). Some translations contain additional annotations which are not part of the original SMS, possibly added by the translators to clarify certain issues with the original message. Along with the noise, spelling variants, and fragmented nature of the SMS mes- sages, the annotations contribute to the overall diffi- culty in building a machine translation system with this type of data. We aim to address some of these issues in out effort. Another challenge with building a Haitian Creole- English translation system is the lack of parallel data. As Haitian Creole is a less commonly spo- ken language, the available resources are limited. Other than the manually translated SMS messages, the available Haitian Creole-English parallel data is about 2 million tokens, which is considerably smaller than the parallel data available for the Stan- dard Translation Task of the WMT11. Lewis (2010) details the effort quickly put forth by the Microsoft Translator team in building a Haitian Creole-English translation system from scratch, as part of the relief effort in Haiti. We took a similar approach to this shared task: rapidly build- ing a translation system to a new language pair uti- lizing available resources. Within a short span (of about one week), we built a baseline translation sys- tem, identified the problems with the system, and exploited …

Content truncated to 500 words