Hi -

The moses recaser appears to essentially be a word-for-word MT model,  
with nothing to keep it from reordering.  And, in fact, I have found  
that it does so, in about 10% of my test sentences.  Here's an  
example, before and after recasing. Note that it swaps "power" and  
"electric".

   the results and based on the quantum 阱 power electric 耦合  
characteristics
   of the concept of pressure is the doctrine of research results are  
consistent .

   The results and based on the quantum 阱 electric power 耦合  
characteristics
   of the concept of pressure is the doctrine of research results are  
consistent . Is this to be expected?

In this case, it's arguably an improvement (although the phrase  
"electric power" does not, in fact, occur in the reference  
sentence.)  In many cases, however, it's doing the wrong thing for my  
data - if I score case-insensitively, BLEU actually goes down by a  
small amount.

I found it surprising that the recaser would do more than what its  
name suggests.  Is this to be expected, and how could I prevent it?   
The recaser model does not appear to include a reordering table, but  
the moses.ini file does have a single weight for reoerdering, 0.6.   
Could I effectively turn off reordering by setting this to 0?

(This may also be related to a recent question by Scott Olsson:  
"preventing reorderings with GIZA++".)

Thanks.

- John Burger
   MITRE

_______________________________________________
Moses-support mailing list
[email protected]
http://mailman.mit.edu/mailman/listinfo/moses-support

Reply via email to