Dear Support, after having read "Factored Models" from Moses's homepage and the corresponding Koehn&Hoang 2007 EMNLP article I'm still uncertain about something, i really hope someone can clarify it for me. Namely I'm not sure how mapping steps with overlapping output factors are handled. For example: "--translation-factors 0-0+1-0" have the output factor 0 overlapping in the two independent translation steps. Or "--translation-factors 0,1-0+2-1 --generation-factors 1-0" have the output factor 0 overlapping between one translation and one generation steps.
So, efficiency concerns for specific mapping combinations aside, how is it handled? I assume it's not via overwriting and ignoring what the previous step suggested. Or are they treated like probabilities and taken independently, so in case of my first example it's log_prob_translation-1 (f0 | f0) * log_prob_translation-2 (f0 | f1)? But in that case the decoding-steps order needn't be specified, since both generation and translation are computed independently and order doesn't matter. Thanks in advance to anyone who clarifies this riddle for me. Best regards, Mark Fishel University of Tartu _______________________________________________ Moses-support mailing list [email protected] http://mailman.mit.edu/mailman/listinfo/moses-support
