Skip to content

Thomas1205/VI3

master
Switch branches/tags
Code

Latest commit

 

Git stats

Files

Permalink
Failed to load latest commit information.
This is a patch for GIZA++, which adds the functionality to compute IBM-3
Viterbi alignments in EM-training.


To install this software, you first need to acquire and compile the
following packages: 

- GIZA++ http://code.google.com/p/giza-pp/
- CBC (tested with Version 2.4.0) http://www.coin-or.org/projects/Cbc.xml

Afterwards, copy the files contained in this patch to the directory where GIZA++ is 
installed. The files are based on the 2009 release of GIZA++ and contain some extra 
code for the computation of IBM-3 Viterbi alignments.

Finally, you will need to adjust the Makefile so that the program is linked to the CBC binaries.
You will need to specify the include directory of CBC and will probably have to add 
-DHAVE_CONFIG_H to the compiler flags to correctly compile the headers of CBC.

Finally, by default the Viterbi alignments are only compared against
the hillclimbing alignments, and some statistics are printed on the
screen. If you want your training to be actually centered around these
alignments, add -DTRAIN_VITERBI to your compiler flags.


Tested under 64-bit Suse Linux with g++ 4.4.2 and 4.5.0.

NOTE: since the publication of the CoNLL 2012 paper, this package has
evolved substantially. In particular, the prior reasoning stage was
substantially improved.  

Also, we note that in the CoNLL 2012 paper we did not respect GIZA++'s
fertility limit. So the comparison to hillclimbing is unfair, but
otherwise the results are correct. If you feel that correct results
should be published, please e-mail the author. This was so far skipped
for lack of time.

About

Computation of IBM-3 Viterbi Alignments (patch for GIZA++)

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages