Philipp Koehn schreef op de 17e dag van de lentemaand van het jaar 2014: > Hi, > > by default, the maximum number of factors is 4, but you are using 5 > output factors. > > You can recompile Moses with an additional switch: > bjam --max-factors=5
I tried that. I get the same error. Hmm, tried that again, and now it does work without error. Strange. > I am a bit concerned about your model since it generates a large number > of output factors independently, which may lead to an explosion of possible > translation options where all good choices get pruned out before you get > to the generation step which will eliminate all the impossible combinations. > > It would be probably a good idea to use this complex factored setup only > as a back-off for unknown words. For details, please look at the following > paper: http://amta2012.amtaweb.org/AMTA2012Files/papers/147.pdf Thanks. I will look into that. I am still trying to understand how to use factors to get a better translator. I have two goals. 1. I want to know what Moses is capable of, to compare its best result to the result of a new translation system that is still being develloped. For this, I need to get an understanding of factored models and hierarchical models (First result with hierarchical model: it works, but worse than basic model). 2. I want to get word alignments that are as good as possible, because we need it for building the new system. And I have seen some very bad word alignments produced by Giza++. Currently, Moses uses factors only in the model. I am wondering if you could use factors to get better word alignments. Another thing I was thinking about is if you could use a trained (and tuned) model to guide the word alignment, that way improving on the original alignment, possably in an iterative process. At the moment, I am still strugling with goal 1. -- Peter Kleiweg http://pkleiweg.home.xs4all.nl/ _______________________________________________ Moses-support mailing list Moses-support@mit.edu http://mailman.mit.edu/mailman/listinfo/moses-support
