This aligner stores the raw, unweighted alignment lattices in a .far archive during training. Currently it is not possible to load these and restart training from this stage. This would be a nice feature, as it would add flexibility and make it possible to train several different models, e.g. with different weight annealing strategies, and/or different EM maximization strategies, while saving the time needed to build the raw lattices. May be useful for large lexica in the 1M+ entry range.
This aligner stores the raw, unweighted alignment lattices in a .far archive during training. Currently it is not possible to load these and restart training from this stage. This would be a nice feature, as it would add flexibility and make it possible to train several different models, e.g. with different weight annealing strategies, and/or different EM maximization strategies, while saving the time needed to build the raw lattices. May be useful for large lexica in the 1M+ entry range.