There are 2 reasons why the tree-to-string model produces translation
even for a string input:
1. the glue rule concatenate all partially translated phrases
together to produce a complete translation. Even for an unknown word, a
translation is available
2. There is small bug in the in-memory implementation of the rule
table. The rule table doesn't check the left-hand-side of the source.
For example, when using the rule
VP --> a ||| X --> b
the VP does not have to agree with the label in the input sentence. This
bug was fixed for the on-disk implementation.
hope that makes sense
On 07/03/2012 16:01, dongxinghua0213 wrote:
hi,
As the name suggests, for tree-to-string model ,the input side
should be a parsed sentence for the training set ,dev set and test set
, e,g,, in English-to-Chinese translation, the source side seems as
follows:
<tree label="NP"> <tree label="DET"> *the* </tree> <tree label="NN">
*cat* </tree> </tree>
but I found the decoder works for the parsed sentences and
unannotated sentences (string) in test set, but the results is
slightly different. Could sombody give some explanations?
_______________________________________________
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
