Elixir/HTML dump scraper

Project link ((in progress)):
https://gitlab.com/wmde/technical-wishes/scrape-wiki-html-dump

A new and wondrous data source has become available to Wikimedia researchers and hobbyists: semantic HTML dumps of all articles on wiki. Previously, archived wiki content was only available as raw wikitext, which is notoriously difficult to parse for information, or even to render. MediaWiki's wikitext depends on a number of extensions and is usually tied to a specific site's configuration and user-generated templates. This recursively parsed content is effectively impossible to expand exactly as it appeared at the time it was written, once the templates and software it depends on have drifted.

HTML dumps are an improvement in every way: content, structure and information are expanded, frozen and made available in a form that can be read by ordinary tools—and the original wikitext is still available as RDFa annotations which makes the new format something like a superset.

At my day job doing Technical Wishes for Wikimedia Germany, we found a reason to dive into these new dumps: it's the only reliable way to count the footnotes on each article. I'll go into some detail about why other data sources wouldn't have sufficed, and also why we're counting footnotes in the first place.

Reference parsing

What are references?

References are the little footnotes all over Wikipedia articles: These footnotes ground the writing in sources, and are a key aspect of the intellectual culture of Wikipedia since all claims are supposed to be paraphrased from existing secondary sources.

Why are we counting references?

The Wikimedia Germany Technical Wishes team has taken the past few months to focus on how references are reused on wikis. We have some ideas about what needs to be fixed (unfortunately this project is currently on hold), but first we needed to take measurements of the baseline situation in order to better understand how references are used, reused, and so that we can tell if our interventions are beneficial.

Previous research into citations has also measured the HTML-formatted articles, but HTML dumps weren't available at the time so this was accomplished by downloading articles one at a time.

For those interested in the results, please skip ahead to the raw summary results, but much more detail and analysis will be published in the future.

Obstacles to finding references in wikitext

A raw reference is straightforward in wikitext and looks like: <ref>This footnote.</ref>. If this were the end of the story, it would be simple to parse references. What makes it more complicated is that many references are produced using reusable templates, for example: {{sfn|Hacker|Grimwood|2011|p=290}}.

If "{{sfn}}" were the only template then we could search for "ref" tags and "sfn" templates in wikitext. But a search for reference-producing templates unveils over 12,000 different templates on English Wikipedia alone, and these will be different on every wiki and language edition.

References are simple in HTML

Once the wikitext is fully rendered to HTML, we can finally see all of the footnotes which were produced. They appear something like this, <div typeof="mw:Extension/ref">Footnote text.</div>

Since the rendering is complete, we know exactly which references are visible, which is better than the potential references that we might have been able to determine from a static analysis of each template.

Template expansion also maps to HTML hierarchical structure, which makes it possible to tell when a reference was produced by templates or when a reference contains templates. Both of these cases are interesting to our research.

Resumability

The scraping job is extremely slow—our first run took two months. If the job crashes for any reason, it's crucial that we can resume again at roughly the same place it was stopped.

We've implemented two levels of resumability and idempotence:

The processing is broken down into small units which each write a single file. If a file already exists, then we skip the corresponding calculation. This general caching technique is known as memoization.

During each step of processing, we also write to a separate checkpoint file after every 100 articles. The output file is written in chunks immediately after writing to the checkpoint file, which reduces the window of time in which a crash results in inconsistency. When resuming, if a checkpoint file is present then the job will skip articles until it catches up with the number given there. The overall behavior of the job is therefore "at least once" processing, meaning that the only type of irregularity that can happen is that some articles might be duplicated in the output.

Concurrency

There are hundreds of separate wikis, so splitting up the work by wiki and processing these concurrently is a natural first implementation.

When splitting by wiki, we ran into an interesting problem where the partitioning function was using :erlang.phash2 to hash an object which contained the wiki ID so we assumed that it would give different results for each wiki, but as it turns out the Flow.partition function needed explicit clues to correctly split by wiki.

The next obvious fork point would be in the phase which makes external API requests, but this is trickier because we want to limit total concurrency across all wikis as well, to avoid overwhelming the service. This should be implemented with a connection pool, ideally one which reuses a small number of connections according to HTTP/1.1 .

Modularity

It will be no surprise that the analyses are run as separate units under a pluggable architecture, so that the tool can be reused for various tasks. The callbacks are crude for now and the abstraction is leaky, but it at least accomplishes code encapsulation and easily configurable composition.

Modules must be written in Elixir, but we're also considering a language-agnostic callback if the need arises.

Some aspects of modularity were no fun so we ignored them. For example, each metric is documented in one big flat file. You'll also encounter some mild hardcoding such as an entire extra processing phase to make external API requests for parsed map data.