Abstract

Pharaoh is a widely-used state-of-the-art decoder for phrasal statistical machine translation. In this paper, we present two modifications to the algorithm used by Pharaoh that together permit much faster decoding without losing translation quality as measured by BLEU score. The first modification improves the estimated translation model score used by Pharaoh to evaluate partial hypotheses, by incorporating an estimate of the distortion penalty to be incurred in translating the rest of the sentence. The second modification uses early pruning of possible next-phrase translations to cut down the overall size of the search space. These modifications enable decoding speed-ups of an order of magnitude or more, with no reduction in the BLEU score of the resulting translations.

‚Äč