all the generation step does is create the word factor, given the brown
cluster. If you only have 50 clusters, and say, all your verbs go into 1
custers, then the generation step would just create lots of verbs,
regardless of the source word.
Not only will this be a bad translation, you will prob run into
time/memory issues when trying to decode
On 29/07/2015 16:26, Marco Damonte wrote:
Thank you! What do you mean by "mapping is too general"? Would you
suggest using the generation for coarse semantic classes (Supersenses)?
Marco
On Wed, 29 Jul 2015 1:21 pm Hieu Hoang <[email protected]
<mailto:[email protected]>> wrote:
This is my setup that generates target words from stems
input-factors = word
output-factors = word stem suffix
alignment-factors = "word -> word"
translation-factors = "word -> stem"
reordering-factors = "word -> word"
generation-factors = "stem -> word"
decoding-steps = "t0, g0"
you're not likely to get anything useful generating words from
brown clusters - the mapping is just too general
On 28/07/2015 20:46, Marco Damonte wrote:
Hi,
I'm using EMS and factorized translation. For instance, I have
this setting:
input-factors = word
output-factors = word brown50
alignment-factors = "word -> word"
translation-factors = "word -> word+brown50"
generation-factors = "brown50"
reordering-factors = "word -> word"
decoding-steps = "t0"
that is, I have brown clusters as an output factor.
Does someone can point me the proper way to add a generation step
to create the surface word using the cluster factor?
I tried:
generation-factors = "brown50 -> word"
but the experiment crashed durinh tuning
Thanks in advance
Marco
_______________________________________________
Moses-support mailing list
[email protected] <mailto:[email protected]>
http://mailman.mit.edu/mailman/listinfo/moses-support
--
Hieu Hoang
Researcher
New York University, Abu Dhabi
http://www.hoang.co.uk/hieu
--
Hieu Hoang
Researcher
New York University, Abu Dhabi
http://www.hoang.co.uk/hieu
_______________________________________________
Moses-support mailing list
[email protected]
http://mailman.mit.edu/mailman/listinfo/moses-support
