Efficient Beam Thresholding for Statistical Machine Translation

Beam thresholding is a widely-used pruning approach in decoding algorithms of statistical machine translation. In this paper, we propose two variations on the conventional beam thresholding, both of which speed up the decoding without degrading BLEU score. The first variation is the dynamic beam thresholding, in which the beam threshold varies with the length of source sequences covered by hypotheses. The second one incorporates a language model look-ahead probability into the beam thresholding so that the interaction between a hypothesis and the contexts outside the hypothesis can be captured. Both thresholding methods achieve significant speed improvements when used separately. By combining them together, we obtain a further speedup, which is comparable to that of the cube pruning approach (Chiang, 2007). Experiments also display that the dynamic beam thresholding can further improve the cube pruning.