Data Mining Wikipedia Notes

I spent quite a bit of time mining Wikipedia while at Qwiki. Our original product was a visual search engine where we used Wikipedia as a main data source to generate short videos from each Wikipedia article. We had to extensively parse the wikitext and associated media from Wikipedia to generate these videos. Here’s an example of our technology being used by Bing:

Qwiki being used in Bing search results.

Qwiki being used in Bing search results.

A Qwiki video of New York City

A Qwiki video of New York City

As a first step to mining Wikipeia for your project, I recommend having a goal in mind. Don’t go down the rabbit hole if you need to just take a peek. Wikipedia is almost 100% open and you can see into their inner workings very easily. But it can be easy to get lost in everything, especially the datasets. Most of the time, you will only need a subset of the data for your goal.

For example, if you want to just grab the article internal links, you can just download the pagelinks MySql table dump and avoid parsing the every article.

There are a few categories I’ll cover:

  • Data Collection
  • Data Filtering
  • Data Extraction

Data Collection

Data Dumps

Wikipedia data comes in two sets: XML Dumps and MySql dumps. The article and revision text are in the XML format and the remaining supplementary data comes in MySql dumps. This includes image metadata, imagelinks, etc.

Both can be found on

You can download the entire article set enwiki-latest-pages-articles.xml or in partitioned chunks: enwiki-latest-pages-articlesX.xml.bz2 where X is 1 to 32

There is also a corresponding rss dump. If you need to be notified when a new wikipedia dump is available, you can write a script to monitor this RSS url. I’ve noticed the dumps commonly released is around the first week of each month.

Besides Wikipedia, there are related datasets which are derived from Wikipedia. The two most popular ones are Google’s Freebase and the open-source DBpedia. Both projects have primary goals of structuring data. By structuring the data, I’m referring to associating a fact with its label. ie. “Bill Gates’ Date of Birth: October 28, 1955″

Freebase comes in an easier to parse format and API. They also release a WEX formatted article text every two weeks. WEX is an xml formatted version of wikitext. Unfortunately, this WEX parser has not seem much updates lately and allegedly doesn’t parse correctly.

BEWARE of using the structured data from Freebase. Some of their data is good but sometimes, the data is very out of date and missing compared to the same wikipedia article. In particular, I noticed their population statistics are out of date. There is also less coverage with location data than Wikipedia.

DBpedia is an open source project written primary in Scala. Their main goal is to extract structured data from the infoboxes. They have a DBpedia Live system where wikipedia articles are automatically updated in DBpedia. Thus DBpedia has near real time updates from wikipedia. You can query for structured data using SPARQL. Using SPAQRL is a little more difficult than Freebase’s JSON-based query language, MQL. They also have supporting projects like DBpedia Spotlight which is designed to extract Wikipedia-derived entities from a body of text.

I would recommend seeing if either of these projects can solve your problem before trying to mine Wikipedia yourself.

Even if you do decide to mine Wikipedia, be sure to use or read the dbpedia parser. They have a lot of information regarding infobox normalization which can help in this area. Consider this page which shows a mapping of the infobox settlement. This can help you think of the heristics you need for your parser.

Data Filtering

Wikipedia has 11 namespaces. If you just want the articles, you can filter against namespace key=’0′. This can greatly reduce the amount of pages you need to process.

Wikipedia also releases just the abstract text in a separate data dump. This is considerably easier to work with if you just need the first few paragraphs of each article.

Data Extraction

The raw Wikipedia data dump comes in XML with the main body text in wikitext. It would have been great if Wikipedia continued to release their article dataset in HTML format but this no longer happens.

Wikitext syntax is very hard to parse and no software has been able to 100% match the PHP mediawiki output. Don’t be too concerned though because in most cases, you won’t need to be 100% perfect.

You will also need to deal with the surrounding metadata stored in the MySql dumps. Take a look at this guide of the mediawiki architecture. It’ll be help to use to decide which data you need.


If you are using a JVM language like Java or Scala, I highly recommend Sweble. Beyond just doing a good job with parsing wikitext, it is a well designed package and it is easy to customize and build upon.

Wikimedia is working on a new parser called Parsoid written in node.js and C++. It is planned to be near compatible with the PHP wikitext parser. It was not as complete when I started mining wikitext so I don’t have experience with it.

What is so problematic about the wikitext format? There are many edge cases and only the original PHP parser has been able to reproduce wikitext to HTML correctly. The other big problem is template expansion which we’ll cover in the next section.

While you can spider the wikipedia website itself for text, Wikimedia recommends you use their data dumps to avoid overloading their servers unnecessarily. I would go as far as cloning wikipedia and running a mirror to get the evaluated html. The biggest problem I’ve encountered with cloning Wikipeia is that each article can take a long time to render. Wikipedia’s production site is heavily cached so reading a page from will be even faster than rendering a local copy.


Beyond the syntactical parsing, the biggest challenge will be how to handle the template expansions. Wikitext has macros called templates which is essentially a language in itself. Templates in wikitext are embedded in {{template_code}}. These can be simple replacement templates or more complex ones with if loops and references to other remote data sources.

Sweble had a template expansion system but I found it didn’t work on edge cases. I resorted to modifying Sweble to call out to the Wikipedia API to expand certain templates.


Beyond just the article text, you may want information with the images and link stats. You’ll need to import the MySql tables to read this information. Here’s a diagram of the mediawiki database schema.

I highly recommend turning off indexes while importing the tables and turning them back on once everything is imported.

Large Scale Processing

You need to determine how much you are processing through the wiki dataset. If you can get away with iterating through the XML dump on a single machine, I highly recommend this approach.

For my purposes, I had to run through a few iterations of the XML dump. Some of the iterations I was able to get away with running it on one machine but other times, I had to parallelize it across multiple machines.

I used Hadoop to perform the parallel processing. Hadoop does have a built-in XML splitter. Mahout also comes with a Wikipedia iterator. I found both of these non-intuitive and incorrect on some cases. I resorted to a system where we collapsed each wiki article xml entry into one line each. Hadoop makes it very easy to parallel process one entry per line datasets.

Image Processing:

Images from Wikipedia articles have two sources: wikipedia itself and commonswiki.

If you are processing images from Wikipedia or Wiki Commons, be aware some images are really large and this can kill your image processing application.

Look at this image metadata, a 26280×19877 98M image…

{"height": 19877, "width": 26280,
"source": ",_by_Jos?_de_Ribera,_from_Prado_in_Google_Earth.jpg",
"url": "" }

The metadata for each image can be found in the MySql database dump table named image. Unfortunately the description field is truncated and you’ll need to join it with the page and revision tables to get the whole description. This description is also in wikitext format so you’ll need to run it through a wikitext parser.

Other Resources

Below are some resources I tried briefly.

Bliki engine: a Java wikipedia parser.

dbpedia’s extraction_framework can be used to extract infoboxes.

GWT Wiki Dump support.

Mahout’s Wikipedia Dump Splitter

Happy Wiki Mining!

Hope this article helps you get started with wikipedia mining!

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s