MRwiki

Tools for analyzing Wikipedia content (especially links) using Python and MapReduce.

Getting started

Below is a sample session in which we download and process Wikipedia dumps using MRwiki tools. Run the following commands in the MRwiki directory.

mkdir tmp
cd tmp
./s1-download-dumps.sh
for f in * ; do gzip -d f ; done
cd ..

At this point the necessary Wikipedia dump files are downloaded and unpacked. It is time to copy them to an HDFS filesystem. Note: we wanted to unpack the dumps, since gzip files are not splittable. If you really care about space, recompress the dump files with bzip2 (not shown here). Warning: do not rename the dump files in any other way than by adding a suffix, since some scripts depend on information that is encoded in the filenames.

export WIKI=hdfs:///user/$USER/wiki
hadoop fs -mkdir $WIKI
hadoop fs -mkdir $WIKI/d1
hadoop fs -put tmp/* $WIKI/d1

Great! The SQL dumps are now on the HDFS. Time to parse them and sort by type. Note: we will use 90 reducers. If you're lucky enough to have a larger cluster at your disposal, go ahead and increase this number. However, bear in mind that allocating more than a thousand or so reducers will probably be a waste of resources.

export COMMONOPTS='-r hadoop --jobconf mapred.reduce.tasks=100 --no-output'
python s2-dumps-to-json.py $WIKI/d1 --output-dir $WIKI/d2 $COMMONOPTS
python s3-select-relations.py $WIKI/d2 --output-dir $WIKI/d3p --rel-type p $COMMONOPTS
python s3-select-relations.py $WIKI/d2 --output-dir $WIKI/d3r --rel-type r $COMMONOPTS
python s3-select-relations.py $WIKI/d2 --output-dir $WIKI/d3ll --rel-type ll $COMMONOPTS
python s3-select-relations.py $WIKI/d2 --output-dir $WIKI/d3pl --rel-type pl $COMMONOPTS

Now we have four directories: d3p with page titles, d3r with raw redirects, d3ll with raw langlinks, and d3pl with raw pagelinks. By "raw" we mean that the links are from pages referenced by identifiers to pages referenced by titles. We don't like it, it's better to have links with identifiers on both sides. Also, we are interested only in pages from the 0th namespace, i.e., in articles.

python s4-match-links.py $WIKI/d3p $WIKI/d3r --output-dir $WIKI/d4r --rel-type r $COMMONOPTS
python s4-match-links.py $WIKI/d3p $WIKI/d3ll --output-dir $WIKI/d4ll --rel-type ll $COMMONOPTS
python s4-match-links.py $WIKI/d3p $WIKI/d3pl --output-dir $WIKI/d4pl --rel-type pl $COMMONOPTS

Excellent! Now all our links are represented in the form: fromId relType toId. Let's focus on redirects for a moment. At this point we most probably have double redirects, triple redirects, even circular ones! Let's clean that up: relink all redirects to the final pages and remove all circular redirects.

python s5-manage-dbl-redirects.py $WIKI/d4r --output-dir $WIKI/d5t1 --shorten 1 $COMMONOPTS
python s5-manage-dbl-redirects.py $WIKI/d5t1 --output-dir $WIKI/d5t2 --shorten 1 $COMMONOPTS
python s5-manage-dbl-redirects.py $WIKI/d5t2 --output-dir $WIKI/d5t3 --shorten 1 $COMMONOPTS
python s5-manage-dbl-redirects.py $WIKI/d5t3 --output-dir $WIKI/d5t4 --shorten 1 $COMMONOPTS
python s5-manage-dbl-redirects.py $WIKI/d5t4 --output-dir $WIKI/d5r $COMMONOPTS

Redirects are now reliable. Let's use them to perform other sanity checks. For example, if a redirect and a different link (lagnlink, pagelink) originates from a page, we remove all the links but the redirect. If a langlink connects two pages from the same langage edition, we remove it.

python s6-check-links.py $WIKI/d5r $WIKI/d4ll --output-dir $WIKI/d6ll $COMMONOPTS
python s6-check-links.py $WIKI/d5r $WIKI/d4pl --output-dir $WIKI/d6pl $COMMONOPTS

Occasionally we will prefer to know where a given link effectively leads to. E.g. instead of a langlink to a redirect to a page, we will prefer a langlink to the final page. Let's calculate such effective links.

python s7-effective-links.py $WIKI/d5r $WIKI/d6ll --output-dir $WIKI/d7ll --rel-type ll $COMMONOPTS
python s7-effective-links.py $WIKI/d5r $WIKI/d6pl --output-dir $WIKI/d7pl --rel-type pl $COMMONOPTS

Technically, langlinks are directed, but intended to be symmetric. Let us perform symmetric closure on langlinks.

python s8-symmetric-links.py $WIKI/d7ll --output-dir $WIKI/d8ll $COMMONOPTS

Let us now try to estimate the credibility of individal langlinks. Here's the idea: we're not sure whether the langlink en:France — fr:France is correct or not, so we look at their surroundings: pagelinks originating from en:France and from fr:France (e.g. en:Paris, en:English channel, fr:Paris, fr:Manche (mer)), and langlinks between targets of these pagelinks. We will say that langlinks like en:Paris — fr:Paris and en:English channel — fr:Manche (mer) are supporting en:France — fr:France. Here is how we calculate the number of supporting langlinks for each langlink.

python s9-find-support.py $WIKI/d6pl $WIKI/d8ll --output-dir $WIKI/d9llSpl $COMMONOPTS

Now, we want to group pages (and langlinks) into components. We will assign to every page and langlink the UUID of component.

python s10.py $WIKI/d3p $WIKI/d8ll --output-dir $WIKI/d10 $COMMONOPTS

Let's check which pages in particular language are connected into one component.

python s11.py $WIKI/d10 --output-dir $WIKI/d11 $COMMONOPTS