Hi Philipp, I will look into tree-to-String and String-to-Tree models. However i did work with relax parses, and used left-binarization and right binarization on my parsed data before I ran training. Although it improved scores over the normal tree-to-tree yet its was still not sufficient to beat the PBSMT. However I am yet to run experiments with SAMT 1-4.
Thanks and regards, Pratyush On 09/04/11 15:53, Philipp Koehn wrote: > Hi, > > the syntax models are still pretty much research-grade and may > disappoint when using them out of the box. > > Especially, tree-to-tree models will lead to much more restricted > models, with less rules, and rules that are less applicable. > > You may see better results by starting with tree-to-string or > string-to-tree models. You should also look at the parse relaxation > which is inspired by work by Zollmann&Venugopal and the > tree binarization work at ISI. > > -phi > > On Fri, Apr 8, 2011 at 11:27 AM, Pratyush Banerjee > <[email protected]> wrote: > >> Hi All, >> >> We have been trying to use Syntax Models for Moses for some time now. >> We have trained tree-to-tree models and HPB models and were trying to >> compare the results with standard PBSMT models. >> >> We use en-de as language pairs and about 1.2 million lines as training data. >> For tree to tree models we have used Berkeley Parser for parsing both >> languages. >> >> However i found that Tree-to-Tree scores were much lower(about 4 BLEU >> points) compared to the PBSMT models. The HPB model (hierarchical moses >> without syntax) is slightly better than the PBSMT models. >> >> Is this behaviour normal. I am quite new to syntax based models, hence the >> question. >> >> Also do we need to parse the devsets during tuning of the tree-to-tree >> models ? >> >> Thanks and regards, >> >> Pratyush Banerjee >> >> _______________________________________________ >> Moses-support mailing list >> [email protected] >> http://mailman.mit.edu/mailman/listinfo/moses-support >> >> >> > _______________________________________________ Moses-support mailing list [email protected] http://mailman.mit.edu/mailman/listinfo/moses-support
