Using GIZA++

If you have parallel corpora you can use GIZA++ to make bilingual dictionaries.

Download your corpora, and convert into one sentence per line.

Download and compile GIZA++.

Use plain2snt.out to convert your corpus into GIZA++ format:

$ plain2snt.out sv-text.txt da-text.txt 
w1:sv-text w2:da-text
sv-text -> sv-text
da-text -> da-text

You may get some warnings about empty sentences like these:

WARNING: filtered out empty sentence (source: sv-text.txt 23 target: da-text.txt 0).
WARNING: filtered out empty sentence (source: sv-text.txt 34 target: da-text.txt 0).

if it is a large corpus you may get a lot of warnings...

After you've done this, you should have a couple of .snt files and a couple of .vcb files.

Next you need to generate word classes, using mkcls:

$ mkcls -m2 -psv-text.txt -c50 -Vsv-text.vcb.classes opt >& mkcls1.log
$ mkcls -m2 -pda-text.txt -c50 -Vda-text.vcb.classes opt >& mkcls1.log

Now use GIZA++ to build your dictionary (-S is the source language, -T is the target language, -C is the generated aligned text file, and -o is the output file prefix):

$ GIZA++ -S sv-text.vcb -T da-text.vcb -C sv-text_da-text.snt -p0 0.98 -o dictionary >& dictionary.log

and wait... You can watch the log in dictionary.log... but the training is likely to take upwards of 10 hours, so have something else planned.

External links

• Moses Language Model Howto v2
• statmt.org: Factored training