Hi, thanks for the reply!
We eventually decided to integrate the feature function in an ad-hoc manner. We didn't find a way to fit our feature in a general interface -- it's stateless but context-sensitive, it needs to get all translations of a phrase at once and it has to be pre-computed before decoding. Best, Ales Dne 29.6.2012 04:29, Barry Haddow napsal(a): > Hi Ales > > Sure, the feature function interface could do with some improvements. It's on > the TODO list, so please go ahead! > http://www.statmt.org/moses/?n=Moses.GetInvolved > >>> Also, is there a reason why scores by stateless features are computed >>> during decoding? If they were computed for each translation option ahead >>> of time (once full translation options are created), the decoder could >>> use them in the initial pruning. It could also prevent repeated >>> computation of feature function values (e.g. our feature is costly to >>> compute). > Yes, features should be precomputed if it improves performance. This is done > with phrase features, but at the moment there is no way to specify that a > given feature should be pre-computed. It's something that needs added. > > cheers - Barry _______________________________________________ Moses-support mailing list [email protected] http://mailman.mit.edu/mailman/listinfo/moses-support
