Christian Hardmeier wrote:
> Hi Linh,
>
>   
>> In some cases, when the training data is very large, the decoder output 
>> *-inf*  in nbest list score ( as in the 1st entry)
>>  or *nan *score (as in the 2nd entry)
>>     
>
> Are you using the latest version of moses? A problem related to infinite
> lexical distortion scores, which is what you seem to have, was fixed a
> few weeks ago. They have to do with probabilities in the lexical
> reordering table getting rounded down to zero by the training script.
> Are there any zero scores in your reordering table?
>
> If you use binary reordering tables produced by an old version of
> processLexicalTable from tables containing zero scores, you'll also
> have to rebuild them with a more recent version of the tool.
>
> Cheers,
> Christian
>
>   

hi Christian,
yeah there are zero scores:

TMAl TMXY5amiyn ||| the secretary - ||| 0.41977 0.00000 0.58022 0.96554 
0.00001 0.03445
TMXY5amiyn ||| secretary - ||| 0.89894 0.00267 0.09839 0.95679 0.00000 
0.04321

so that is the problem, perhaps  change 0.00000 --> 0.000001 can fix 
this problem
thanks so much

linh

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

Reply via email to