Reranking for Large-Scale Statistical Machine Translation
Kenji Yamada, Ion Muslea · The MIT Press eBooks · 2008
Statistical machine translation (SMT) systems, which are trained on parallel corpora of bilingual text (e.g., French and English), typically work as follows: for each sentence to be translated, they generate a plethora of possible translations, from which they keep a smaller n-best list of the most likely translations. Even though the typical n-best list contains mostly high-quality candidates, the actual ranking is far from accurate. This chapter presents a novel approach to reranking the n-best list produced by an SMT system. It uses an ensemble of perceptrons that are trained in parallel, each of them on just a fraction of the available data. Experiments were performed on two large-scale commercial systems: a Chinese-to-English system trained on 80 million words and a French-to-English system trained on 1.1 billion words. The reranker obtained statistically significant improvements of about 0.5 and 0.2 BLEU points on the Chinese-to-English and the French-to-English system, respectively.