PORTAGE: with Smoothed Phrase Tables and Segment Choice Models

Proceedings of the Workshop on Statistical Machine Translation, pages 134–137, New York City, June 2006. c⃝2006 Association for Computational Linguistics PORTAGE: with Smoothed Phrase Tables and Segment Choice Models Howard Johnson National Research Council Institute for Information Technology Interactive Information 1200 Montreal Road Ottawa, ON, Canada K1A 0R6 Howard.Johnson@cnrc-nrc.gc.ca Fatiha Sadat, George Foster, Roland Kuhn, Michel Simard, Eric Joanis and Samuel Larkin National Research Council Institute for Information Technology Interactive Language Technologies 101 St-Jean-Bosco Street Gatineau, QC, Canada K1A 0R6 firstname.lastname@cnrc-nrc.gc.ca Abstract Improvements to Portage and its partici- pation in the shared task of NAACL 2006 Workshop on Statistical Machine Trans- lation are described. Promising ideas in phrase table smoothing and global dis- tortion using feature-rich models are dis- cussed as well as numerous improvements in the software base. 1 Introduction The statistical machine translation system Portage is participating in the NAACL 2006 Workshop on Sta- tistical Machine Translation. This is a good opportu- nity to do benchmarking against a publicly available data set and explore the benefits of a number of re- cently added features. Section 2 describes the changes that have been made to Portage in the past year that affect the par- ticipation in the 2006 shared task. Section 3 outlines the methods employed for this task and extensions of it. In Section 4 the results are summarized in tab- ular form. Following these, there is a conclusions section that highlights what can be gleaned of value from these results. 2 Portage Because this is the second participation of Portage in such a shared task, a description of the base system can be found elsewhere (Sadat et al, 2005). Briefly, Portage is a research vehicle and development pro- totype system exploiting the state-of-the-art in sta- tistical machine translation (SMT). It uses a custom built decoder followed by a rescoring module that adjusts weights based on a number of features de- fined on the source sentence. We will devote space to discussing changes made since the 2005 shared task. 2.1 Phrase-Table Smoothing Phrase-based SMT relies on conditional distribu- tions p(s|t) and p(t|s) that are derived from the joint frequencies c(s, t) of source/target phrase pairs ob- served in an aligned parallel corpus. Traditionally, relative-frequency estimation is used to derive con- ditional distributions, ie p(s|t) = c(s, t)/ � s c(s, t). However, relative-frequency estimation has the well-known problem of favouring rare events. For instance, any phrase pair whose constituents occur only once in the corpus will be assigned a probabil- ity of 1, almost certainly higher than the probabili- ties of pairs for which much more evidence exists. During translation, rare pairs can directly compete with overlapping frequent pairs, so overestimating their probabilities can significantly degrade perfor- mance. To address this problem, we implemented two simple smoothing strategies. The first is based on the Good-Turing technique as described in (Church and Gale, 1991). This replaces each observed joint frequency c with cg = (c + 1)nc+1/nc, where nc is the number of distinct pairs with frequency c (smoothed …

Content truncated to 500 words