Hi,
Short answer: add --discount_fallback to the command line (and if the
option isn't defined, update the code).
Technically the closed-form estimate for Kneser-Ney discounts isn't
defined for models that have no singletons, such as class-based models.
The --discount_fallback option allows you to make up discounts. You
probably don't know which ones to use, but they don't really matter for
unigrams because 10000 - 1.5 is about the same as 10000 anyway. So just
use the default which is 1.5 for counts 3 and above.
Kenneth
On 10/11/14 08:45, CHATZITHEODOROU Konstantinos wrote:
> Hi all,
>
> I am trying to train a factored phrased model using KenLM but I get an
> error message during pos_train LM.
>
> Please see attached files for more details.
>
> Thanks in advance for your time and your consideration.
>
> Best,
>
> CHATZITHEODOROU Konstantinos
>
>
>
> _______________________________________________
> 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
