- creates a virtual phrase table at decoding time based on a vector of component models and a combination algorithm
- linear interpolation or instance weighting
- two possible component model types supported so far: 0 (in-memory) or 12 (compact)
- weights can be set in config, and overriden on a sentence-level through mosesserver API
- online optimization (perplexity minimization) using dlib and xmlrpc-c call