Lattice-based Consensus decoding for Moses

People: Abhishek Arun, Barry Haddow

The implementation of MBR decoding in Moses currently uses n-best lists. Recent work has shown the benefits of using the decoder's lattice for improved and faster consensus decoding: e.g lattice MBR (Tromble et al. 2008, Kumar et al. 2009) or linear approximations of lattice MBR (De Nero et al. 2009, Pauls et al. 2009) In this project, we propose to implement these algorithms in Moses. To do so, will require adding a framework for processing the Moses output search graph (lattice). Some of the operations in this framework: a) Implementing the forward-backward algorithm b) Fast extraction of all n-grams in the lattice c) Rescoring the lattice d) Computing the rescored lattice best path This project should be doable in a week's time or close to it.

Page last modified on January 25, 2010, at 03:11 PM