Improved Minimum Error Rate Training in Moses

We describe an open-source implementation of minimum error rate training (MERT) for statistical machine translation (SMT). This was implemented within the Moses toolkit, although it is essentially standsalone, with the aim of replacing the existing implementation with a cleaner, more flexible design, in order to facilitate further research in weight optimisation. A description of the design is given, as well as experiments to compare performance with the previous implementation and to demonstrate extensibility.