Philipp Koehn wrote:

> if the input is provided as a lattice, then it is not
> the case that all translations have the same number
> of unknown words.

Moreover, the penalty is certainly needed at weight-training time, correct?  
Then there are different input sentences with different amounts of OOV.  The 
way I anthropomorphize this is that it helps the weights "explain" why some 
sentences are harder to translate than others.

- John Burger
  MITRE

> On Mon, Oct 3, 2011 at 11:28 AM, Kaveh Taghipour
> <[email protected]> wrote:
>> Hi Christian,
>> 
>> Thanks. You are right. But I think there is no need for such a penalty,
>> since all candidates for a given source sentence contain the same number of
>> OOVs and so the penalty does not help at all. Do you know the reason?
>> 
>> Cheers,
>> Kaveh
>> 
>> On Sun, Oct 2, 2011 at 9:53 PM, Christian Hardmeier <[email protected]> wrote:
>>> 
>>> I think you're on the right track. For some reason, moses doesn't report
>>> the OOV penalty feature, which adds a hardcoded penalty of -100 to the total
>>> score for each input word that was copied to the output because no suitable
>>> translation was found in the phrase table. Your test sentence probably
>>> contains two unknown words that account for the -200 difference between the
>>> score you calculated and the one output by the decoder. Does that make
>>> sense?
>>> 
>>> Best,
>>> Christian
>>> 
>>> Kaveh Taghipour <[email protected]> wrote:
>>> 
>>>> Hi,
>>>> 
>>>> I have generated an N-best list with moses, but I do not know how to
>>>> compute
>>>> the final score. I tried a log-linear model but came up with a wrong
>>>> number.
>>>> For example:
>>>> 
>>>> moses.ini:
>>>> ----------------------------------------
>>>> # distortion (reordering) weight
>>>> [weight-d]
>>>> 0.120662
>>>> 
>>>> # language model weights
>>>> [weight-l]
>>>> 0.251853
>>>> 
>>>> # translation model weights
>>>> [weight-t]
>>>> 0.0675335
>>>> 0.103843
>>>> 0.0720954
>>>> 0.00630806
>>>> 0.101934
>>>> 
>>>> # word penalty
>>>> [weight-w]
>>>> -0.275771
>>>> 
>>>> Assuming a log-linear model, the score for the following sentence should
>>>> be
>>>> *-54.076* but the decoder prints:
>>>> 
>>>> 0 ||| ....  ||| d: 0 lm: -219.367 w: -6 tm: -3.17999 -4.43847 -2.63049
>>>> -3.87513 3.99959 ||| *-254.076*
>>>> 
>>>> Thank you in advance for helping.
>>>> 
>>>> Cheers,
>>>> Kaveh
>>>> 
>>>> _______________________________________________
>>>> 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

Attachment: smime.p7s
Description: S/MIME cryptographic signature

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

Reply via email to