Abstract

Modern SMT systems incorporate multiple components, statistical models, and processes. Translation is often factored as a series of modules with the output of one module serving as the input to the next. To avoid propagation of errors, it is better to avoid hard decisions by passing on as much information as possible to subsequent stages of the MT pipeline, usually in the form of a lattice or list of the most likely hypotheses. This enables the application of models that are difficult or impossible to apply in first-pass translation.

I will describe several large-scale SMT lattice rescoring procedures based on minimum Bayes-risk decoding, starting with an efficient implementation of lattice MBR that uses weighted path counting transducers to compute the required statistics. This implementation allows efficient generalisation of the MBR decoder to the task of multiple-lattice system combination. I will conclude by describing a confidence-based lattice segmentation and MBR decoding framework; this framework enables the targeted application of models intended to address particular deficiencies in SMT hypotheses.