A Beam Search Algorithm for ITG Word Alignment.

Peng Li,Yang Liu,Maosong Sun

2012-01-01

Abstract:Inversion transduction grammar (ITG) provides a syntactically motivated solution to modeling the distortion of words between two languages. Although the Viterbi ITG alignments can be found in polynomial time using a bilingual parsing algorithm, the computational complexity is still too high to handle real-world data, especially for long sentences. Alternatively, we propose a simple and effective beam search algorithm. The algorithm starts with an empty alignment and keeps adding single promising links as early as possible until the model probability does not increase. Experiments on Chinese-English data show that our algorithm is one order of magnitude faster than the bilingual parsing algorithm with bitext cell pruning without loss in alignment and translation quality.