[Corpora-List] [Moses-support] how to increase reordering diversity of n-best list generated by Moses

Philipp Koehn pkoehn at inf.ed.ac.uk
Fri Oct 10 20:28:05 CEST 2014


Hi,

to obtain an n-best list with reordering diversity, you could dump out the search graph and search for reorderedly diverse translations according to a your preferences.

-phi

On Fri, Oct 10, 2014 at 9:11 AM, Saeed Farzi <saeedfarzi at gmail.com> wrote:


> Dear All,
>
>
> I am going to increase reordering diversity of n-best list generated by
> Moses in order to apply my reordering model (a re-ranker system) for
> improving translation quality.
>
> In the first step due to low reordering diversity of the n-best list, i
> need that the n-best list is included by different sentences which have
> different reorderings.
>
>
> For increasing the n-best list diversity, we generate 10000-best list and
> then select 100-best of sentences with different reorderings.
>
> For generating 10000-best list we use stack diversity and cube-pruning
> options.
> There are two disadvantages:
>
> 1) This task is very time consuming because of generating 10000-best in
> the first step.
>
> 2) we loose lots of good translation candidates in terms of
> lexical choice. It is caused that our reranker system could not find better
> translation candidates in order to improve translation quality (BLEU score).
>
>
> I am looking for new solution for increasing reordering diversity
> without aforementioned problems.
>
> Cheers
>
>
> --
> S.Farzi, Ph.D. Student
> Natural Language Processing Lab,
> School of Electrical and Computer Eng.,
> Tehran University
> Tel: +9821-6111-9719
>
>
>
> _______________________________________________
> Moses-support mailing list
> Moses-support at mit.edu
> http://mailman.mit.edu/mailman/listinfo/moses-support
>
>
-------------- next part -------------- A non-text attachment was scrubbed... Name: not available Type: text/html Size: 2681 bytes Desc: not available URL: <https://mailman.uib.no/public/corpora/attachments/20141010/92ec785a/attachment.txt>



More information about the Corpora mailing list