Fast decoding and optimal decoding for machine translation

Abstract

A good decoding algorithm is critical to the success of any statistical machine translation system. The decoder's job is to find the translation that is most likely according to set of previously learned parameters (and a formula for combining them). Since the space of possible translations is extremely large, typical decoding algorithms are only able to examine a portion of it, thus risking to miss good solutions. In this paper, we compare the speed and output quality of a traditional stack-based decoding algorithm with two new decoders: a fast greedy decoder and a slow but optimal decoder that treats decoding as an integer-programming optimization problem.

Generating under Global Constraints: the Case of Scripted Dialogue

“…Particularly relevant here is the work of Germann et al (2001) who explicitly compare optimal algorithms for translation with fast algorithms and use the former to evaluate the latter. dp ap emph aggr cand gen cand gen sol gen time Table 2: Results of running our dialogue revision module on eight different initial dialogue plans.…”

Online Learning via Dynamic Reranking for Computer Assisted Translation

“…Different search strategies have been suggested to define the way in which the search space is organised. Some authors [33,18] have proposed the use of an A algorithm, which adopts a best-first strategy that uses a stack (priority-queue) in order to organise the search space, and this strategy has become the most commonly used. On the other hand, a depth-first strategy was also suggested in [2], using a set of stacks to perform the search.…”

Monte Carlo techniques for phrase-based translation

“…Our sampler is similar to the decoder of (Germann et al, 2001;Eisner and Tromble, 2006;Langlais et al, 2007), which start with an approximate solution and then incrementally improve it via operators such as RETRANS and MERGE-SPLIT. It is also similar to the estimator of Marcu and Wong (2002), who employ the same operators to search the alignment space from a heuristic initialisation.…”