you must tune so that the scores from the decoder correlate to the BLEU scores to get good results.
to create tuning data, just hold out 1000-2000 random sentence pairs from the existing corpora from the training data and use them as your tuning data On 5 April 2010 12:03, nati g <[email protected]> wrote: > Hi, > currently i had configured moses on rhel 5.3 and languge is english to > italian. > > i have't tuned the translation model as i am not sure of creating tuning > data set from existing corpus. > i would be great helps if you can direct me in creating a tuning data set > from existing corpus. > > Thanks in advance. > > > On Mon, Apr 5, 2010 at 3:41 PM, Hieu Hoang <[email protected]> wrote: > >> hi nati >> >> you haven't given much detail of your system or the problems you see in >> the translation. which languages are you translating? >> >> i assume you have run done all the usual things - train a translation and >> language model and tuned with MERT. >> >> the easiest thing to do is collect more data for the language model, >> especially data in the same field as the sentences you are translating. The >> language model is monolingual so the data is easier to obtain. Scrape the >> web, word documents, UN charter, anything you can find. If you have LOTS of >> LM data, try increasing the LM order >> >> hieu >> >> >> On 05/04/2010 09:54, nati g wrote: >> >> Hi, >> I had tried an translation model from a corpus of 1.5 million lines. but >> the out translation quality is not that good. >> >> i would like to know what all possiblites that we have to, improve the >> translation qualtity while training the translation model. >> >> Thanks in advance. >> >> >> _______________________________________________ >> 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 >> >> >
_______________________________________________ Moses-support mailing list [email protected] http://mailman.mit.edu/mailman/listinfo/moses-support
