You can use mgiza's force alignment feature. Check out this tutorial. It is
not real-time, however.
http://www.kyloo.net/software/doku.php/mgiza:forcealignment
--Q
On Thu, Oct 9, 2014 at 9:24 PM, Wei-Yun Ma wm2...@columbia.edu wrote:
Dear all,
I have trained a workable Chi-Eng Moses
Unfortunately for Model 3/4 you can't get Viterbi probability - Viterbi
decoding for fertility models is not implemented in giza. Viterbi decoding
for fertility model is exponential. So the prob you get in the final
alignment file is an approximation using hill climbing.
If you want to do
MGIZA++ can be compiled using VC++ both 32 and 64bit, however 64bit version
occasionally crashes during final clean up. So YMMV. Also, I believe
mingw-w64 is out there, maybe you can try that (msys + mingw-w64)
http://stackoverflow.com/questions/9942923/mingw-as-a-reliable-64-bit-gcc-compiler
Model 3/4 uses HMM/Model2 to bootstrap the alignment. First Viterbi
alignment is computed using Model2 or HMM model, and then a hillclimbing
algorithm is used to find optimal alignment using model 3/4. So even the
initial alignment has some problem you _may_ still get good alignment
result. So I
On 20 January 2012 14:17, Qin Gao q...@cs.cmu.edu wrote:
I may missed early mails in this thread, are you using giza++ or mgiza?
--Q
On Fri, Jan 20, 2012 at 9:11 AM, Guchun Zhang gzh...@alphacrc.comwrote:
Hi again,
I got the config file done as suggested by the sample config file.
However
These are lexical translation and distortion models produced by previous
training, and can be produced by removing -nodump and -onlyaldumps from
giza options in moses training scripts.
For documentation of giza and mgiza parameters, you can refer to
Yes, MGIZA isn't really incrementally training, it only initialize the
model parameters with that trained previously, since it does not store
sufficient statistics of the previous training. It will give bad performance
if
1. You train only model 1 or
2. The incremental data or sub set is really
If you comment out -DBINARY_SEARCH_FOR_TTABLE then giza will use hash table
instead of binary search for lexicon translation model, that will use more
memory. The problem seems to be snt2cooc.out is not installed. My suggestion
is to keep the compiler flag and install snt2cooc.out, make sure that
0.999 -s
/home/amin/mt/work/baseline/phraseBased/3lm/training/prepared.1/en.vcb -t
/home/amin/mt/work/baseline/phraseBased/3lm/training/prepared.1/fr.vcb
died with signal 6, without coredump
On Tue, Dec 28, 2010 at 8:25 PM, Qin Gao q...@cs.cmu.edu wrote:
If you comment out
No, you just need to run it on multi-core nodes in order to benefit from
multi-threading.
Also it seems to me you compiled giza++ with hashtable-based TTable
implementation, that should not happen if you are using newest version of
giza++ or mgiza++, otherwise you have to check
Hi,
The mkcls and snt2cooc utility in mgiza++ are almost unchanged from GIZA++.
But snt2cooc's command line interface is changed: instead of writing to
STDOUT, it writes to a file. I don't think it can be used without changing
Moses script, and the change is mainly to support Hadoop based
From: moses-support-boun...@mit.edu [moses-support-boun...@mit.edu] On
Behalf Of Qin Gao [q...@cs.cmu.edu]
Sent: Wednesday, December 09, 2009 10:21 PM
To: moses-support@mit.edu
Subject: [Moses-support] -lmodel-dub parameter
Hi List,
Could anyone please
If you run GIZA only up to model 1, the final output will be
src-tgt.A1.*
instead of src-tgt.A3.final
You have to modify the script to manually rename it.
Also, this page may help you on GIZA parameters.
http://geek.kyloo.net/software/doku.php/mgiza:configure
--Q
On Fri, Dec 11, 2009 at
Hi List,
Could anyone please explain what is the usage of -lmodel-dub parameter? Does
it related to language model filtering? In the help message it says
-lmodel-dub parameter but no furthur information is given e.g. format of
dictionary, how the boundary works etc.
Thanks a lot!
--Q
.
___
Moses-support mailing list
Moses-support@mit.edu
http://mailman.mit.edu/mailman/listinfo/moses-support
--
==
Qin Gao
Language Technology Institution
Carnegie Mellon University
http://geek.kyloo.net
Hi,
I wonder whether the word alignment information in the phrase table is used
in decoding? Such as ||| (0) (1,2,3) () ||| (2) (1,2) |||. I think moses
does not use the lexicon, so how does these information be used?
Thanks.
Qin
___
Moses-support
Hi, Wilson,
As I mentioned, GIZA++ may have a bug on HMM training stage, it will add
some random number to count table, and maybe it is the reason. You may
check the archive of the mailing list for the description of the bug,
also, you can simply comment out the lines marked with //***//
, have you been able to determine if your fix
corrects the problem with the German-English alignment?
Chris
On Thu, Feb 28, 2008 at 12:50 PM, Qin Gao [EMAIL PROTECTED] wrote:
Hi, Wilson,
As I mentioned, GIZA++ may have a bug on HMM training stage, it will add
some random number to count
Hi All,
I found a bug in GIZA++ source, which may make slight difference in hmm
training, and if you link GIZA++ with pthread library, it will crash the
training. To keep it short, just do modification to Array2.h, from line
63 to 74:
---
inline T*begin(){
#ifdef __STL_DEBUG
if(
19 matches
Mail list logo
