On 14 June 2013 18:24, Mohammad Salameh <[email protected]> wrote:

> Thank You German and Hie for your reply
> It is just from what I noticed that the transition score of a hypothesis
> is it is score minus the score of the hypothesis stated in the "Back"
> field.(am I right? )
>
> Also I tried multiplying the core scores with the weights from moses.ini
> (and then sum over the multiplication), but I did not get the the right
> score indicated by the hypothesis above. How can I know the correct
> ordering of the core scores so that I achieve correct multiplication ?
>

at the beginning, there should be lines like this:
   FeatureFunction: KENLM0 start: 0 end: 1
   FeatureFunction: Distortion0 start: 1 end: 2
   FeatureFunction: WordPenalty0 start: 2 end: 3
   ...
that is the order of the score vector

> Thanks,
> Salameh
>
>
>
>
> On Fri, Jun 14, 2013 at 2:23 AM, Hieu Hoang <[email protected]> wrote:
>
>> German is right, the extra score is for the feature
>>    UnknownWordPenalty
>> It's usually not shown because it's weight doesn't need to be tuned.
>>
>> To 'unhide' it, put this in your ini file
>>   UnknownWordPenalty tuneable=true
>>
>> On 14 June 2013 09:55, Germán Sanchis Trilles <[email protected]>wrote:
>>
>>> Hi,
>>>
>>> there are 15 core scores because one of them (probably one of the zeroes
>>> in the example you provided) is the out-of-vocabulary penalty, which is
>>> zero if there is none, or -100 for each OoV.
>>>
>>> The "score", afaik, is the accumulated score, while the "transition" is
>>> the combination of the "core" scores, which you should be able to achieve
>>> by multiplying the "core" scores by the weights in the moses.ini, taking
>>> into account that the order has to match and that there is a OoV penalty (0
>>> in this case) with weight 1.
>>>
>>> Cheers,
>>>
>>> Germán
>>>
>>>
>>>
>>>
>>>
>>> On Thu, 13 Jun 2013, Mohammad Salameh wrote:
>>>
>>>  Hi I am working on the osgx file and I want to make sure what the core
>>>> scores in bold
>>>> below means.
>>>>
>>>> 0 hyp=92 stack=1 back=0 score=-1.622 transition=-1.622 forward=2346
>>>> fscore=-7.220
>>>> covered=4-4scores="core=0.000,**-1.000,-1.705,-1.946,-4.290,-**
>>>> 3.850,1.000,0.000,0.000,-0.**336,0.000,0
>>>> .000,0.000,-4.000,-15.454) " out="reject|AlrAfDp"
>>>>
>>>>
>>>> Each hypothesis has 15 core features while the moses,ini below has 14
>>>> features. Why is
>>>> there an extra score?
>>>>
>>>>
>>>> LexicalReordering0= 0.0763367 0.00720199 0.0806019 0.0411903 0.0565713
>>>> 0.0805696
>>>> Distortion0= 0.0400278
>>>> LM0= 0.0903804
>>>> WordPenalty0= -0.168358
>>>> TranslationModel0= 0.0558195 0.0296544 0.0225311 0.0427546 0.208002
>>>>
>>>> Also, are the core scores used to compute the hypothesis score=-1.622
>>>> the above ?
>>>>
>>>> Regards,
>>>> Salameh
>>>>
>>>>
>>>>
>>> _______________________________________________
>>> Moses-support mailing list
>>> [email protected]
>>> http://mailman.mit.edu/mailman/listinfo/moses-support
>>>
>>>
>>
>>
>> --
>> Hieu Hoang
>> Research Associate
>> University of Edinburgh
>> http://www.hoang.co.uk/hieu
>>
>>
>


-- 
Hieu Hoang
Research Associate
University of Edinburgh
http://www.hoang.co.uk/hieu
_______________________________________________
Moses-support mailing list
[email protected]
http://mailman.mit.edu/mailman/listinfo/moses-support

Reply via email to