Hi Nghia
You mean the phrase-based decoder, or chart decoder? The phrase-based
hypothesis recombination is follows this paper
www.ldc.upenn.edu/acl/W/W01/W01-1408.pdf
It's slightly better than just looking @ the trailing words because it
takes into account the backoff structure of the LM
On 17/06/2010 07:45, [email protected] wrote:
Dear Hieu Hoang,
I am a PhD student and I am trying to understand what Moses does.
Currently, I have some confusions that I would like to seek clarification
from you.
I tried to print out some trace from Moses to see how it recombines and
drop weaker paths. However, what I observed was not consistent with what I
read from the tutorial in www.stamt.org.
The thing is that sometimes, Moses recombines hypotheses which only match
in the coverage vector but does not match in the trailing words (consider
the last two trailing words - I am using 3-gram model).
For example, when I input the English word "in" and run the decoder, I see
that all but one hypothesis are recombined despite the fact that their
translations do not share the last two words.
It is very important to me to truly understand what Moses does. Thus, it
would be a great appreciate if you could help me to deconfuse.
Thank you very much.
Regards
Hoang Trong Nghia
_______________________________________________
Moses-support mailing list
[email protected]
http://mailman.mit.edu/mailman/listinfo/moses-support
