Hi Barry, I have observed the following things regarding the feature weights initialisation.
The previous declaration of weights in the "moses.ini" is overridden by the weights initialisation done in the "moses-mert-new.pl". Please correct me if I am wrong. I have the following doubt. 1. What is the difference between word penalty and phrase penalty (used as one of five feature weights of translation) ? On Wed, Jan 14, 2009 at 4:16 PM, Barry Haddow <[email protected]> wrote: > Hi Taraka > > I'll try to respond as best as I can inline. > > On Tuesday 13 January 2009 22:07, you wrote: > > RDear Barry, > > > > I have the following doubts with the mert implementation. > > > > 1. What is decoder flags switch meant for in mert syntax? > These are the flags that are passed through to the decoder, for example if > you > want to specify the stack size. > > > 2. When you wrote statistics in Bleu score, I went through the code > and > > didnt quite understand it. What is the meaning of it? > The purpose of the Scorer::prepareStatistics() is to perform any > pre-calculation required so that the calculation of the metric during > optimisation is as fast as possible. In the case of bleu, this means > counting > the number of n-gram matches and calculating the length penalty. > > > 3. Infact I am using the mert implementaion for letter to phoneme > > conversion system. Thats why the values are odd. > Does bleu make sense in this domain? I guess you don't have any reordering? Bleu doesnt make much sense in this domain. Still with the default weights the system is achieving comparable results for L2P task. There is no reordering in this. I want to remove all reordering and penalties. > > > > 4. Your suggestion of "make clean all" helped a lot. I was able to run > > mert with LENGTH values of 1,3,4,6. I observed that other than 1 and 6 > > the system simply crashed. > Could you be more specific. Saying "the system simply crashed" doesn't help > me. > > > 5. I want to keep the weights of word penalty and distortion model to > > zero and activate only translation model weights and language model > > weights. How can I do that? > Just remove them from your model. Take the penalty out of the phrase table > and > update moses.ini accordingly. > > > 6. I suppose keeping the weights zero means the value of model is > always > > 1, taking exponential of zero = 1. > That's correct. If the weight is zero, then the feature has no effect. > > > 7. Is it possible to have any other loss function like mWER? > It should be. Just create a new Scorer class. > > > 8. I thought that only the forward translation (log)phrase translation > > probabilities are taken into account for translation model. Why are > > backward probabilities taken into account for a feature function? > Because they improve translation? I'm afraid I can't give a better answer, > but > you could try asking this on the moses list, > > regards > Barry > > -- > The University of Edinburgh is a charitable body, registered in > Scotland, with registration number SC005336. > > -- With Regards, Taraka Rama.
_______________________________________________ Moses-support mailing list [email protected] http://mailman.mit.edu/mailman/listinfo/moses-support
