Thanks Ondrej. I think I understood the difference between the two phases but I'm still wondering which weights the decoder assigns if the translation option attached to a hypothesis does not appear on the reordering table.
Suppose you have two phrases "a1 | b1" and "a1 | b2", but your reordering table only has weights for "a1 | b1". What happens then? Thank you again. Regards. ----- Mensaje original ---- De: Ondrej Bojar <[email protected]> Para: Carlos Henriquez <[email protected]>; [email protected] Enviado: mié,21 julio, 2010 16:23 Asunto: Re: [Moses-support] Using two translation tables AND two reordering tables Dear Carlos Henriquez, the reordering model is applied only once the translation options are attached to a hypothesis to extend it. The two (alternative) decoding paths are used only during the first phase: construction of translation options. In other words, one reordering model covers both decoding paths. On the contrary, I'm not sure whether Moses at all supports scoring partial hypotheses with two reordering tables, regardless the decoding steps that were used to construct the translation options. Cheers, Ondrej. "Carlos Henriquez" <[email protected]> wrote: > >________________________________ > >Hi all. > >I want to build a SMT system with two translation tables. I "know" how to >specify that in the moses.ini file but I have doubts about the reordering model. > >Suppose I built those different translation tables from two corpora. Also, >both >corpora generated a reordering table. > >I saw in Moses/Advance Feature's webpage >[http://www.statmt.org/moses/?n=Moses.AdvancedFeatures#ntoc14] that I can tell >the decoder to have two decoding path but I don't see how to handle the >different reordering table. > >Could I do something like this? > >[ttable-file] >0 0 5 /my-dir/table1 >0 0 5 /my-dir/table2 > >[distortion-file] >0 0 5 /my-dir/d-table1 >0 0 5 /my-dir/d-table2 > >[mapping] >0 T 0 >1 T 1 > >so Moses will know that decoding path 0 goes with table1 and d-table1 only and >path 1 goes with the other two. > > >On the other hand, if I use only one reordering table. How does Moses solve >the >phrases that do not have reordering weights? > > > > > >_______________________________________________ >Moses-support mailing list >[email protected] >http://mailman.mit.edu/mailman/listinfo/moses-support -- Ondrej Bojar http://www.cuni.cz/~obo _______________________________________________ Moses-support mailing list [email protected] http://mailman.mit.edu/mailman/listinfo/moses-support
