Can I use incremental GIZA++ for the new lines, even though I didn't use it for the baseline? (does mgiza give me everything inc-giza needs?)
If not, I like the idea of just running word alignment on the new lines. Would I need to update any files besides *.A3.final.gz for steps 3+ to run correctly? (do steps 3+ use any previously computed files aside from these?) Elliot On Fri, Jul 26, 2013 at 11:30 AM, Philipp Koehn <pko...@inf.ed.ac.uk> wrote: > Hi, > > you could just run word alignment on the 50,000 lines, but you will get > better performance if you somehow leverage the baseline parallel corpus > for word alignment. > > One way is incremental GIZA++, the other is re-run everything. > > You could also try some middle ground of including some of the baseline > data in a re-running word alignment. > > It is not clear how much you will loose by going down these options... > > -phi > > On Fri, Jul 26, 2013 at 2:16 AM, Elliot K Meyerson > <ekmeyer...@wesleyan.edu> wrote: > > Hello, > > > > I have a large phrase-based translation system. Alignment was done with > > mgiza, and took a few weeks. I now have a small amount of extremely > relevant > > new bitext (~50,000 lines) that I would like to use to augment the model, > > without having to retrain everything. The new data contains many > important > > words that are not found anywhere else in the training data, so lexical > > tables (at least) would need to be updated along with adding in new > > alignments. I could run the rest of training (steps 3+) no problem, as > long > > as the relevant files from steps 1 and 2 are updated in a reasonable > way. Is > > there some way for me to do this? or should I just cut my losses and > retrain > > the entire thing? > > > > Thanks, > > Elliot > > > > _______________________________________________ > > Moses-support mailing list > > Moses-support@mit.edu > > http://mailman.mit.edu/mailman/listinfo/moses-support > > >
_______________________________________________ Moses-support mailing list Moses-support@mit.edu http://mailman.mit.edu/mailman/listinfo/moses-support