Myndighetscrawl

Myndighetscrawl is small project to gather and analyse PDF files on Swedish government websites, using several open sources for web crawling such as the Internet Archive and Common Crawl.

The problem

The Swedish government consists of a great number of agencies (statliga förvaltningsmyndigheter) that publish several thousands of analyses and reports every year. Unfortunately, their decentralised nature and the lack of best open data practices makes them hard to find, and even harder to gather as a consolidated and structured dataset.

Some agencies consistently publish their reports on a specific "Publications" page, but many do not. Some add metadata to the PDF files, many do not. In practice, the work of automatically fetching documents from hundreds of websites is too much and most organisations do not keep their old reports online.

These publications are also archived by three organisations:

• the Swedish National Library (Kungliga biblioteket), still partly in paper form
• the Swedish National Archives (Riksarkivet), still partly in paper form and often with a delay of several years
• the Government Offices of Sweden (Regeringskansliet) in their digital registry (diariesystem) for those documents sent to the government

They have obligations to give access to these documents through the old Swedish access to information law (offentlighetsprincipen) but as of today, none of them release the data to the public.

How does Myndighetscrawl try to solve this problem?

Myndighetscrawl looks through open sources that crawl the Internet such as the Internet Archive and Common Crawl. It uses their APIs to extract lists of PDF files that were archived from government websites by the services.

The source code is still work in progress and can be found on Github.

What data is available right now?

The data made available is the raw list of links provided by the services.

Internet Archive

The data is in archive_org.parquet.

The Internet Archive provides:

• timestamp, a timestamp for the time the document was archived
• original, the URL at which the document was found
• length, the size in bytes of the archived file

The file also contains two fields generated by the data:

• archive, the URL to access the file on the Wayback Machine, made with timestamp and original
• filename, the filename extracted from the URL if possible

Common Crawl

The data is in common_crawl.parquet.

The Common Crawl Index API provides:

• Domain, the base domain at which the file was fetched
• Period, a monthly or yearly reference to when it was fetched
• URL, the URL at which the document was found

What's next?

The two sources contain respectively 2.16 and 1.25 million rows. The next steps are to:

• identify the documents that hold a high value, like reports
• filter out duplicates
• determine important metadata about the documents such as their title, register number and potentially other connections with important processes (government missions...)

This can't only be done by looking at the URL of the document, they need to be downloaded and analysed. A script to do it using ML and LLM is in preparation.

If you want to try and analyse the documents in this dataset, don't hesitate to reach out so we can join forces!