NRC‘s PORTAGE System for WMT 2007

Proceedings of the Second Workshop on Statistical Machine Translation, pages 185–188, Prague, June 2007. c⃝2007 Association for Computational Linguistics NRC’s PORTAGE system for WMT 2007 Nicola Ueffing, Michel Simard, Samuel Larkin Interactive Language Technologies Group National Research Council Canada Gatineau, Qu´ebec, Canada firstname.lastname@nrc.gc.ca Howard Johnson Interactive Information Group National Research Council Canada Ottawa, Ontario, Canada Howard.Johnson@nrc.gc.ca Abstract We present the PORTAGE statistical machine translation system which par- ticipated in the shared task of the ACL 2007 Second Workshop on Statistical Machine Translation. The focus of this description is on improvements which were incorporated into the system over the last year. These include adapted lan- guage models, phrase table pruning, an IBM1-based decoder feature, and rescor- ing with posterior probabilities. 1 Introduction The statistical machine translation (SMT) sys- tem PORTAGE was developed at the National Research Council Canada and has recently been made available to Canadian universities and research institutions. It is a state-of-the-art phrase-based SMT system. We will shortly de- scribe its basics in this paper and then high- light the new methods which we incorporated since our participation in the WMT 2006 shared task. These include new scoring methods for phrase pairs, pruning of phrase tables based on significance, a higher-order language model, adapted language models, and several new de- coder and rescoring models. PORTAGE was also used in a joint system developed in coop- eration with Systran. The interested reader is referred to (Simard et al., 2007). Throughout this paper, let sJ 1 := s1 . . . sJ de- note a source sentence of length J, tI 1 := t1 . . . tI a target sentence of length I, and ˜s and ˜t phrases in source and target language, respectively. 2 Baseline As baseline for our experiments, we used a ver- sion of PORTAGE corresponding to its state at the time of the WMT 2006 shared task. We pro- vide a basic description of this system here; for more details see (Johnson et al., 2006). PORTAGE implements a two-stage transla- tion process: First, the decoder generates N- best lists, using a basic set of models which are then rescored with additional models in a sec- ond step. In the baseline system, the decoder uses the following models (or feature functions): • one or several phrase table(s), which model the translation direction p(˜s | ˜t). They are generated from the training corpus via the “diag-and” method (Koehn et al., 2003) and smoothed using Kneser-Ney smooth- ing (Foster et al., 2006), • one or several n-gram language model(s) trained with the SRILM toolkit (Stolcke, 2002); in the baseline experiments reported here, we used a trigram model, • a distortion model which assigns a penalty based on the number of source words which are skipped when generating a new target phrase, • a word penalty. These different models are combined log- linearly. Their weights are optimized w.r.t. BLEU score using the algorithm de- scribed in (Och, 2003). This is done on the provided development corpus. The search algorithm implemented …

Content truncated to 500 words