Wikipedia dumps

Wikipedia dumps are useful for quickly getting a corpus. They are also the best corpora for making your language pair are useful for Wikipedia's Content Translation tool :-)

You download them from

• http://dumps.wikimedia.org/backup-index.html

Tools to turn dumps into plaintext

There are several tools for turning dumps into useful plaintext, e.g.

• Wikipedia Extractor – a python script that tries to remove all formatting
• wikiextractor – another python script that removes all formatting (with different options), putting XML marks just to know when begins and ends everty single article
• mwdump-to-pandoc – shell wrapper around pandoc (see the usage.sh below the script for how to use)
• Calculating_coverage#More_involved_scripts – an ugly shell script that does the job
• wp2txt – some ruby thing (does this work?)

Content Translation dumps

There are also dumps of the articles translated with the Content Translation tool, which uses Apertium (and other MT engines) under the hood:

• https://www.mediawiki.org/wiki/Content_translation/Published_translations

To turn a tmx into a "SOURCE\tMT\tGOLD" tsv, `sudo apt install xmlstarlet` and do:

$ zcat ~/Nedlastingar/cx-corpora.nb2nn.text.tmx.gz          \
  | xmlstarlet  sel -t -m '//tu[tuv/prop/text()="mt"]'      \
       -c 'tuv[./prop/text()="source"]/seg/text()' -o $'\t' \
       -c 'tuv[./prop/text()="mt"]/seg/text()'     -o $'\t' \
       -c 'tuv[./prop/text()="user"]/seg/text()'   -n       \
  > nb2nn.tsv

Now view the word diff between MT and GOLD with:

$ diff -U0 <(cut -f2 nb2nn.tsv) <(cut -f3 nb2nn.tsv) | dwdiff -c | less

and find the original if you need it in cut -f1 nb2nn.tsv