Incrementally Updating the SMT Reordering Model

PACLIC 28 !272 Incrementally Updating the SMT Reordering Model Shachar Mirkin Xerox Research Centre Europe 6 Chemin de Maupertuis, Meylan, France shachar.mirkin@xrce.xerox.com Abstract This work is concerned with incrementally training statistical machine translation (SMT) models when new data becomes available. That, in contrast to re-training new models based on the entire accumulated data. In- cremental training provides a way to per- form faster, more frequent model updates, en- abling keeping the SMT system up-to-date with the most recent data. Specifically, we address incrementally updating the reorder- ing model (RM), a component in phrase-based machine translation that models phrase order changes between the source and the target languages, and for which incremental train- ing has not been proposed so far. First, we show that updating the reordering model is helpful for improving translation quality. Sec- ond, we present an algorithm for updating the reordering model within the popular Moses SMT system. Our method produces the exact same model as when training the model from scratch, but doing so much faster. 1 Introduction Parallel data for training statistical machine trans- lation (SMT) models is being constantly generated, both by professional and by casual translators. Typ- ically, large amounts of data are required to produce decent SMT models, yet training a model is an ex- pensive process in terms of time and computational resources. Most often, and in particular when com- munity effort is made to translate new content, it is desirable to keep the system up-to-date with the new data; yet, constant retraining is not feasible. The line of research concerning incremental training for SMT has been addressing this problem, aiming at updating the model given new parallel data, rather than retraining it. Typical phrase-based SMT models use a log- linear combination of various features that mostly represent three sub-models: a translation model (TM), responsible for the selection of a target phrase for each source phrase, a language model (LM), addressing target language fluency, and a reorder- ing model (RM). The reordering model is required since different languages exercise different syntac- tic ordering. For instance, adjectives in English pre- cede the noun, while they typically follow the noun in French (the blue sky vs. le ciel bleu); in Mod- ern Standard Arabic the verb precedes the subject, and in Japanese the verb comes last. As a result, source language phrases cannot be translated and placed in the same order in the generated transla- tion in the target language, but phrase movements have to be considered. This is the role of the re- ordering model. Estimating the exact distance of movement for each phrase is too sparse; therefore, instead, the lexicalized reordering model (Koehn, 2009) estimates phrase movements using only a few reordering types, such as a monotonous order, where the order is preserved, or a swap, when the order of two consecutive source phrases is inverted when their translations are placed in the target side. Most research on incremental training for SMT addresses parallel corpus alignment, the slowest step of the model …

Content truncated to 500 words