Hi all, Currently i would like to exploit the Moses search graph as an alternative for N-best list to improve MT, thanks to a huge number of hypotheses it contains. Nevertheless when i try to output all the hypotheses in the search graph, i observe that they don't cover all N-best list. In other words, there are still several hypotheses in N-best (N=1000) not included in the search graph. Even when i increase the number of hypotheses popped for each stack (up to 10000 for example), this coverage can not be obtained. So, i would like to ask all of you whether we have any Moses option (command, script…) to ensure that the search graph will contain all N best list of each source sentence? For information, the command we invoked to generate the search graph and the N-best list is as follows:
./home/toolkits/Moses/moses-04-2009/moses-cmd/src/moses -f output10881/moses.ini -mbr -cube-pruning-pop-limit 10000 -s 10000 -output-search-graph search_graph_881 -use-alignment-info -print-alignment-info -include-alignment-in-n-best -print-alignment-info-in-n-best -n-best-list ./Nbest-881test 1000 < ./pre-processed-corpus/881.lowercased.fr > ./pre-processed-corpus/881.hyp.SG Furthermore, the way we conduct to extract all hypotheses (sentences) from the search graph is : We start from each final node, move backward until reach the initial one. First, the hypothesis is set : s= "". In each step, we get the field "out" of the current node (the translated word(s)) and add to the hypothesis : s = out + s. Is this the right way to obtain all hypotheses? Thanks a lot and all suggestions would be very welcomed ! bests, Quang _______________________________________________ Moses-support mailing list [email protected] http://mailman.mit.edu/mailman/listinfo/moses-support
