How hard would it be to append the LMBR scores to the list of features instead 
of overwriting it? Maybe I can tackle this at MTMA15 next week. I’m not too 
worried about the long runtime at least initially.

> On May 6, 2015, at 5:01 PM, Barry Haddow <[email protected]> wrote:
> 
> Hi Jeremy
> 
> It's been a long time since I looked at this, but I think these are the 
> component scores in the linearised corpus bleu used in Lattice MBR (see 
> Tromble et al (2008), section 5). They are used to rescore the nbest list, 
> and must be implemented by replacing the original feature vectors.
> 
> So it looks like lattice MBR doesn't work with nbest lists, or at least 
> doesn't give you what you need for tuning. Do you really want to tune with 
> lattice MBR? It's going to be very slow,
> 
> cheers - Barry
> 
> On 06/05/15 20:49, Jeremy Gwinnup wrote:
>> Hi,
>> 
>> I’ve been attempting to experiment with lattice MBR with various settings 
>> and I see something weird happen to my nbest lists:
>> 
>> 0 ||| the prime ministers of india and japan meet in tokyo ||| Distortion0= 
>> … etc etc becomes
>> 
>> 0 ||| the prime ministers of india and japan meet in tokyo ||| map: 0 w: 12 
>> 11.19 8.06 6.63 5.39
>> 
>> My feature weights get replaced by map and w feature with 5 weights
>> 
>> I’m setting the following as moses.ini parameters:
>> lminimum-bayes-risk
>> lmbr-p
>> lmbr-r
>> mbr-scale
>> lmbr-pruning-factor
>> 
>> Anything extra that I need to know as far as what map and w are, or how to 
>> get my normal features back in the nbest list?
>> 
>> Thanks!
>> -Jeremy
>> _______________________________________________
>> Moses-support mailing list
>> [email protected]
>> http://mailman.mit.edu/mailman/listinfo/moses-support
> 
> 
> -- 
> The University of Edinburgh is a charitable body, registered in
> Scotland, with registration number SC005336.
> 


_______________________________________________
Moses-support mailing list
[email protected]
http://mailman.mit.edu/mailman/listinfo/moses-support

Reply via email to