Machine Translation

, Volume 24, Issue 2, pp 103-121

First online:

Monte Carlo techniques for phrase-based translation

  • Abhishek ArunAffiliated withUniversity of Edinburgh Email author 
  • , Barry HaddowAffiliated withUniversity of Edinburgh
  • , Philipp KoehnAffiliated withUniversity of Edinburgh
  • , Adam LopezAffiliated withUniversity of Edinburgh
  • , Chris DyerAffiliated withUniversity of Maryland
  • , Phil BlunsomAffiliated withOxford University Computing Laboratory

Rent the article at a discount

Rent now

* Final gross prices may vary according to local VAT.

Get Access


Recent advances in statistical machine translation have used approximate beam search for NP-complete inference within probabilistic translation models. We present an alternative approach of sampling from the posterior distribution defined by a translation model. We define a novel Gibbs sampler for sampling translations given a source sentence and show that it effectively explores this posterior distribution. In doing so we overcome the limitations of heuristic beam search and obtain theoretically sound solutions to inference problems such as finding the maximum probability translation and minimum risk training and decoding.


Statistical machine translation Gibbs sampling Machine learning MCMC