It should work. The function
EvaluateInIsolation()
in the LM is for optimisation reason. eg. if the target phrase is 'a b c
d' and the LM is a trigram, the trigrams 'a b c' and 'b c d' can be
precalculated in EvaluateInIsolation().
Implementing a pt for factors requires setting up some variables, which
may not have happen yet in mmsapt. if you can send me a small example
model, i'll see what i can do
On 01/09/2015 02:11, Ulrich Germann wrote:
Hi Michael,
I have no experience with factored models, so I'm speculating here to
some degree. The reason the phrase table calls EvaluateInIsolation is
because all "isolated" phrase scores are considered when pruning. In
my opinion pruning should not happen within the phrase tables (for
exactly the reason that it does not allow feature functions to be
agnostic about other feature functions) but by whatever object calls
all the phrase tables and does the generation. However, for software
legacy reasons, that's the way it is right now, and I'm not likely to
address this issue any time soon myself. The most reasonable fix for
this in my opinion is to move pruning where it belongs --- post all
the factor generation stuff.
Hieu is probably still the person with the best understanding of how
factored phrase table entry generation works, so maybe he can chime in
on this ...
Cheers - Uli
On Mon, Aug 31, 2015 at 11:29 PM, Michael Denkowski
<[email protected] <mailto:[email protected]>>
wrote:
Hi Ulrich,
I was looking into using a class-based LM with your dynamic phrase
table via generation models. I translate factor 0 to 0 with the
Mmsapt, then generate target factor 1 (word class) with a GM. The
class-based LM operates on factor 1.
I'm hitting a segfault on what appears to be an
order-of-operations issue with the PT and LM. In mmsapt.cpp:578,
Mmsapt::mkTPhrase makes a call to tp->EvaluateInIsolation. This
calls all of the models, including the LMs. The class LM tries to
score factor 1, which doesn't exist yet (since generation happens
after translation), and it dies. By nature, other phrase tables
don't have this issue since they can just pull up pre-computed scores.
Is scoring with all of the models here a strategic choice to get
better performance or would it be sufficient to just score with
the PT features? Thanks!
--Michael
--
Ulrich Germann
Senior Researcher
School of Informatics
University of Edinburgh
_______________________________________________
Moses-developers mailing list
[email protected]
http://mailman.mit.edu/mailman/listinfo/moses-developers
--
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
