Increasing Translation Speed in Phrase-based Models via Suboptimal Segmentation

Phrase-Based Models constitute nowadays the core of the state of the art in the statistical pattern recognition approach to machine translation. Being able to introduce context information into the translation model, they usually produce translations whose quality is often difficult to improve. However, these models have usually an important drawback: the translation speed they are able to deliver is mostly not sufficient for real-time tasks, and translating a single sentence can sometimes take some minutes. In this paper, we describe a novel technique for reducing significantly the size of the translation table, by performing a Viterbi-style selection of the phrases that constitute the final phrase-table. Even in cases where the pruned phrase table contains only 6% of the segments of the original one, translation quality is not worsened. Furthermore, translation quality remains the same in the worst case, achieving an increase of 0.3 BLEU in the best case,