Output details
11 - Computer Science and Informatics
University of Edinburgh
A Gibbs sampler for phrasal synchronous grammar induction
<22> Originality: We show how the heuristics associated with creating statistical translation grammars (an approach used by virtually all machine translation researchers) can be removed using modern Bayesian non-parametric methods. Additionally we show how Bayesian inference can be made to scale to larger corpora than before.
Significance: Statistical Machine Translation is driven by heuristics. We show how these heuristics can be cleanly removed using a prior over grammars. This more strongly relates translation with machine learning.
Rigour: Uses optimisation on a training set and evaluation on a test set. Paper appeared in the top NLP venue (acceptance rate 21%).