On 08/11/2012 12:27, Felipe Sánchez Martínez wrote:
Hi Hieu,
Thank you very much for your answer.
Or could labels overlap in a way which is not compatible with a single
tree structures? For instance:
<tree label="NP" span="0-2"/>
<tree label="DET" span="0-0"/>
<tree label="N1" span="1-2"/>
<tree label="N2" span="0-1"/>
<tree label="ADJ" span="2-2"/>
la bruja verde
yep. During decoding, it will either of the incompatible labels, but not
both labels.
I understand that during decoding all of the rules learned from the
incompatible spans will be considered, and of course only one finally
used when producing a translation.
Note: the tree-to-* doesn't build trees on the source side, it just uses
the source labels as conditions when apply the translation rules.
Ok, but the application of the translation rules ends up producing
kind of a tree or a set of trees (thanks to the glue rules) when a
complete tree cannot be build. Am I right?
it produces a derivation that looks like a tree. However, the nodes in
the source derivation may not be labelled consistently. For example, if
you have an input tree
(S (NP (PRO he)) (VP slept))
The sentence could be decoded with a grammar:
PRO --> he ||| X --> je
VP --> slept ||| X --> dorm
S --> PRO VP ||| X --> X X
However, it could also be decoded with this grammar:
PRO --> he ||| X --> je
VP --> slept ||| X --> dorm
S --> NP VP ||| X --> X X
In this case, the node above the terminal 'he' is labelled as both NP
and PRO.
So the tree-to-* is probably a slightly incorrect name for the model. It
perhaps should be called source-syntax-constrained model. And the
translation rules should be written like this:
X --> he ||| X --> je , if LHS non-term=PRO
X --> slept ||| X --> dorm , if LHS non-term=VP
X --> X X ||| X --> X X , if LHS non-term=S and 1st non-term=NP
and 2nd non-term=VP
Also explained in Section 4.2.10 and 4.2.11.
Hope that makes some sense
Exact definition of the 'tree'-to-* decoding algorithm is in my big
writeup, section 4.2.8
http://www.statmt.org/~s0565741/ddd.pdf
Thanks for the pointer.
Cheers
--
Felipe
_______________________________________________
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
