On 08.02.2017 13:21, Francis Tyers wrote:
> El 2017-02-08 00:34, Nikita Medyankin escribió:
>> Now it wasn't. I'm not sure whether it was because of the lack of time
>> at the end of the summer or because the gain from using weighted
>> transfer seemed to be negligible. My personal opinion is that the
>> weakest spot (both performance- and quality-wise) is the python module
>> which calculates the coverages and is used in the weight training
>> tool. Regarding this, it seems like an idea to adapt the C++ code from
>> the core to calculate the coverages during training instead of using
>> that python module.
>>
>
> I have a couple of comments to this:
>
> 1) One reason it could seem negligible is that there weren't so many 
> ambiguous
>    transfer rules in the tests. That the performance did not decrease 
> is already
>    a result! :)
Yeah, there were like 3 groups of ambiguous rules with 2 rule variants 
in each, out of some 200 or 300 rules total.)

> 2) Would it be possible to merge the changes to the C++ code ? Would 
> it cause any
>    problems in pairs that don't use the functionality ?
I would need to refresh my memories about it but I believe the stuff 
about the transfer weights was specifically desiged to skip all 
transfer-weight-related code if there's no weights file passed when the 
transfer module is called.

> Fran


------------------------------------------------------------------------------
Check out the vibrant tech community on one of the world's most
engaging tech sites, SlashDot.org! http://sdm.link/slashdot
_______________________________________________
Apertium-stuff mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/apertium-stuff

Reply via email to