MERT tunes the individual feature weights (i.e. 1.1-1.5 in your
example, as well as individual feature weights of the other types).

On Thu, Sep 17, 2009 at 5:37 AM, Lu Wei <[email protected]> wrote:
> Hi,
>
> According to the manual of Moses, the MERT only tunes the weights for the
> following 4 components:
> 1. phrase translation
> 2. language model
> 3. distortion model
> 4. word penalty
>
> where the phrase translation can be decomposed into the following:
> 1.1. P(f|e)
> 1.2. P(e|f)
> 1.3. Pw(f|e)
> 1.4. Pw(e|f)
> 1.5. phrase-penalty: exp(-1)
>
> It seems Moses won't assign a separate weight to each individual components
> 1.1-1.5 above, instead it will assign a single weight to the combined
> component "phrase translation". Am I right? In other words, Moses assumes
> the same weight should get assigned to all the 5 components above? Was the
> same done by Pharaoh also?
>
> --
> best,
> Wei
>
> Sent from Singapore
>
> _______________________________________________
> 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