W dniu 10.02.2014 20:46, Barry Haddow pisze:
Ah, by the way, is removing the Distortion feature from the ini file and 
setting the limit to 1 a safe way to actually disable distortion? Moses 
does not complain (I always thought it is required.)
Best,
Marcin

> Hi Marcin
>
> I had some fun with --activate-features in the past - I think the 
> syntax was rather strange. If it is not working now, it may have got 
> dropped by the recent refactoring
>
> My advice would be to use kbmira (or pro), since they are regularised 
> they don't go crazy when there is an uninformative feature. That way, 
> you don't have to fiddle with feature activation,
>
> cheers - Barry
>
> On 10/02/14 18:01, Marcin Junczys-Dowmunt wrote:
>> Hi,
>> it seems --activate-features=STRING is not working in mert-moses.perl.
>> The script prints a message that the ignored features are not being
>> used, but then optimizes them anyway. I can see that the "enabled"
>> information in the feature data structure is not being used anywhere in
>> the script once it has been set (apart from printing the message).
>>
>> This can cause an interesting catastrophe when, for instance, distortion
>> is disabled by setting the limit to 1:
>> MERT assigns a weight of 1 to distortion (but the feature itself is
>> always 0) and 0 weights to all other features, the final score is then
>> equal to 0 for all sentences and poor moses goes crazy generating lots
>> of garbage which in turn takes ages to score only to finish with bad
>> weights. Really ugly, took me a while to find the cause :)
>>
>> BTW. In my opinon a --deactive-features might be more useful. I would
>> add/correct it myself, but currently I am getting lost in the code that
>> is printing the config files. Someone more acquainted with that code?
>> Best,
>> Marcin
>> _______________________________________________
>> 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

Reply via email to