TC-STAR Speech Translation Workshop, June 20, 2006, Barcelona, Spain
This paper describes the participation of the machine translation team at NRC Canada in the Mandarin-to-English open evaluation task of the 2006 TC-STAR Workshop on Speech-to-Speech Translation. We describe PORTAGE, a statistical phrase-based machine translation system, and discuss the experimental results for two variants of the system obtained in the evaluation. Both variants of the system used hand-coded rules to translate numbers and dates, along with a small number of Chinese names (those encountered in the dev corpus). The secondary variant of PORTAGE carried out Good-Turing smoothing of the phrase tables. Both variants, especially the primary one, performed extremely well on the two metrics related to weighted N-gram models (WNM); we suspect this is due to the incorporation of rules for translating Chinese names.