Wikipedia Ql Save

Query language for efficient data extraction from Wikipedia

Project README

WikipediaQL: querying structured data from Wikipedia

WikipediaQL is an experimental query language and Python library for querying structured data from Wikipedia. It looks like this:

$ wikipedia_ql --page "Guardians of the Galaxy (film)" \
      [email protected] as "title";
      section[heading="Cast"] as "cast" >> {
          li >> text:matches("^(.+?) as (.+?):") >> {
              text-group[group=1] as "actor";
              text-group[group=2] as "character"
      section[heading="Critical response"] >> {
          sentence:contains("Rotten Tomatoes") as "RT ratings" >> {
              text:matches("\d+%") as "percent";
              text:matches("(\d+) (critic|review)") >> text-group[group=1] as "reviews";
              text:matches("[\d.]+/10") as "overall"

RT ratings:
  overall: 7.8/10
  percent: 92%
  reviews: '334'
- actor: Chris Pratt
  character: Peter Quill / Star-Lord
- actor: Zoe Saldaña
  character: Gamora
- actor: Dave Bautista
  character: Drax the Destroyer
- actor: Vin Diesel
  character: Groot
- actor: Bradley Cooper
- actor: Lee Pace
  character: Ronan the Accuser
- actor: Michael Rooker
  character: Yondu Udonta
- actor: Karen Gillan
  character: Nebula
- actor: Djimon Hounsou
  character: Korath
- actor: John C. Reilly
  character: Rhomann Dey
- actor: Glenn Close
  character: Irani Rael
- actor: Benicio del Toro
  character: Taneleer Tivan / The Collector
title: Guardians of the Galaxy (film)


WikipediaQL-the-tool does roughly this:

  • Parses query in WikipediaQL-the-language;
  • Uses MediaWiki API to fetch pages' metadata;
  • Uses Parsoid API to fetch each page's content semantic HTML;
  • Applies selectors from the query to page to extract structured data.

The WikipediaQL development is covered in ongoing series of articles: (newest first)

Subscribe to follow


“Sometimes magic is just someone spending more time on something than anyone else might reasonably expect.” — Raymond Joseph Teller

Wikipedia is the most important open knowledge project: basically, the "table of contents" of all the human data. While it might be incomplete or misleading in details, the amount of data is incredible, and its organization makes all the data accessible to humans.

OTOH, the data is semi-structured and quite hard to extract automatically. This project is an experiment in making this data accessible to machines—or, rather, to humans with programming languages. The main goal is to develop an easy to use and memorize, unambiguous and powerful query language and support it by the reference implementation.

See FAQ below for justifications of parsing Wikipedia instead of just using already formalized Wikidata (and of parsing HTML instead of Wikipedia markup).


$ pip install wikipedia_ql

$ wikipedia_ql --page "Page name" query_text
# or
$ wikipedia_ql query_text_with_page

Usage as Python library:

from wikipedia_ql import media_wiki

wikipedia = media_wiki.Wikipedia()

data = wikipedia.query(query_text)

Full WikipediaQL query looks like this:

from <source> {

When using --page parameter to the executable, you need only to pass selectors in the query text.

Source is Wikipedia article name, or category name, or (in the future) other ways of specifying multiple pages. Selectors are similar to CSS; they are nested in one another with selector { other; selectors }, or (shortcut) selector >> other_selector. All terminal selectors (e.g., doesn't having others nested) produce values in the output; the value can be associated with a name with as "valuename".

See below for a list of selectors and sources currently supported and the future ones.


Simple query for some info from the page:

$ wikipedia_ql --page "Pink Floyd" \
    'section[heading="Discography"] >> li >> {
        a as "title";
        text:matches("\((.+)\)") >> text-group[group=1] as "year";

- title: The Piper at the Gates of Dawn
  year: '1967'
- title: A Saucerful of Secrets
  year: '1968'
- title: More
  year: '1969'
- title: Ummagumma
  year: '1969'
#     ...and so on...

Multi-page query from pages of some category (only from inside Python):

query = r'''
from category:"2020s American time travel television series" {
    [email protected] as "title";
    section[heading="External links"] >> {
      li >> text:matches("^(.+?) at IMDb") >> text-group[group=1] >> [email protected] as "imdb"

# iquery returns generator, fetching pages as you go
for row in wikipedia.iquery(query):
# {'title': 'Agents of S.H.I.E.L.D.', 'imdb': ''}
# {'title': 'The Flash (2014 TV series)', 'imdb': ''}
# {'title': 'Legends of Tomorrow', 'imdb': ''}
# ....

Navigating through pages in one query (note the -> which means "perform subquery in the page by link"):

$ wikipedia_ql --page Björk \
    'section[heading="Discography"] >> li >> a -> {
        [email protected] as "title";
        .infobox-image >> img >> @src as "cover"

- cover:
  title: Debut (Björk album)
- cover:
  title: Post (Björk album)
- cover:
  title: Homogenic

As the page source should be fetched from Wikipedia every time, and it can be a major slowdown when experimenting, wikipedia_ql implements super-naive caching:

wikipedia = media_wiki.Wikipedia(cache_folder='some/folder')
wikipedia.query('from "Pink Floyd" { [email protected] }') # fetches page from Wikipedia, then runs query
wikipedia.query('from "Pink Floyd" { [email protected] }') # gets the cached by prev.request contents from some/folder

(Caution! as it was said, for now, the cache is super-naive: it just stores page contents in the specified folder forever. You might delete it from the cache manually, though: there are just PageName.meta.json and PageName.json files.)

Current state & Planned features

Query language

from <source> {

Source can be:

  • "Page title"
  • category:"Category title"
  • "Page title","Other page title" (several pages at once)
  • geo:"<lat>, <lng>"
  • prefix:"Page title prefix"
  • search:"Search string"
    • Maybe several search types, as Wikipedia has several search APIs

Selectors are CSS-alike selectors, type.class[attr="value"][otherattr="othervalue"]. Note, that unlike CSS, nesting (any child inside parent) is performed not with spaces (parent child), but with parent >> child.

  • Singular selectors:
    • regular CSS selectors, like a or table.wikitable
      • Note again, that selector nested_selector is not supported, so you need to li >> a to say "all links inside the list items"
    • section
      • section[heading="Section heading"]: fetch everything inside section with the specified heading (full heading text must match);
      • section:first (useful to fetch article intro)
      • section: all sections;
      • section[level=3]: all sections of particular level
      • more powerful heading value patterns would be supported (probably in CSS-alike manner: heading^="Starts from" and so on)
    • text
      • text:matches("pattern"): part of the document matching pattern (Python's regexp); document's structure would be preserved, so you can nest CSS and other WikipediaQL selectors inside: li >> text:matches("^(.+?) as") >> [email protected] as "link"
      • text: without pattern specification, just selects the entire text of the parent element;
      • pattern flags, like text:imatches("pattern") (case-insensitive)
      • handle inline images alt attribute as text
    • text-group: should be directly nested in text pattern, refers to capture a group of the regexp; see the first example in the README;
      • text-group[group=1]: group by number
      • text-group[group="name"]: named groups
    • sentence
      • sentence:contains("pattern"): find sentence where pattern matches (whole sentence is selected)
      • sentence: all sentences in the scope
      • sentence:first
      • other CSS pseudo-selectors for sentence
      • pattern flags (same as for text)
      • more suitable sentence tokenizer will be used, probably: currently we are relying on nltk, which is too powerful (and large dependency) for our simplistic needs
    • page: refers (from any scope) to the entire current page; useful for re-nesting fetched data in a logical way and to include metadata attributes in output (see below)
    • <selector>@<attribute>:
      • <css_selector>@<tag_attribute>
        • expand URLs
      • [email protected]<page_attribute>
        • title
        • other
        • smart fetching metadata
      • as a free-standing selector (@title on the top level to fetch "current page's title" instead of [email protected], and so on)
    • table-data for data tables: from "Kharkiv" { section[heading="Climate"] >> table >> table-data >> tr[title^="Average high"] >> td[column="Jan"] }; see docs
    • infoboxes are covered with the same table-data quasi-selector, see showcase
    • hatnote to fetch and process Hatnotes (special links at the top of the page/section, saying "for more information, see [here]", "this page is about X, if you want Y, go [there]" and so on)
    • :has: like CSS :has pseudo-class but support all the WikipediaQL selectors, so one might say from category:"Marvel Cinematic Universe films" { :has(@category*="films") { with pages... } } to drop from the result-set pages of the category which aren't movies.
    • (?) :primary or something like that (maybe :largest), to select the most important thing in the scope (for example, section[heading="Discography"] >> ul:primary will probably fetch the list of albums, while the section might have other, smaller lists, like the enumeration of studios where recordings were done)
    • TBC on "as-good-examples-found" basis
  • Selector operations:
    • sequence: parent >> child
    • group: { selector1; selector2; selector3 } fetch all the selectors in the result set
    • immediate child: parent > child; maybe (come good reasons) other CSS relations like sibling1 + sibling2
    • follow link: section["Discography"] >> li >> a -> { selectors working inside the fetched page }, to allow expressing page navigation in a singular query
  • Marking information to extract:
    • extract unnamed data: every terminal selector puts extracted value in resultset (the resultset then will look like several nested arrays)
    • as "variablename": every terminal selector with associated name puts extracted value as {"name": value}; there is still some uncertainty on how it all should be structured, but mostly the right thing is done
    • as :type and as "name":type for typecasting values
      • (?) simpe types (as "year":int) maybe wouldn't be that necessary, as the conversion can be easily done in the client
      • as :html (as opposed to current "content text only" extraction) might be useful in many cases
      • converting a large section of the document into composite type should shine. Things like infobox as :hash or wikitable as :dataframe will change the usability of data extraction significantly
  • Other query language features:
    • comments (probably // text, as # can be a start of valid CSS selector)
    • (maybe?) named/reusable subquires

Other features/problems

  • Speed is not stellar now (roughly, a few seconds per page). This is due to a) Parsoid API doesn't support batch-fetching pages, so each page is fetched with a separate HTTP request, and b) this is an unoptimized prototype (so parsing takes a lot more than it could). To fix this, we plan to implement:
    • Less naive page caching
    • Profiling and optimizations (like probably using naked lxml instead of BeautifulSoup, and simpler sentence tokenizer)
    • Support for async/parallel processing (as far as I can understand, in Python async I/O would be the most useful optimization; but multi-threaded selectors processing will bring no gain due to GIL?)
  • More robust edge case handling is planned, like links to absent pages, redirects (including redirecting to the middle of another page), disambiguation pages, etc.
  • Exposing HTTP client dependency to the client code, allowing request logging, custom caching strategies, etc.
  • Some potential far-future features:
    • using other languages Wikipedias: nothing in WikipediaQL makes it bound to English Wikipedia only
    • (maybe?) other MediaWiki-based sites (like Wikvoyage, Wiktionary, etc.): more high-level selectors (like infobox and wikitable) would be irrelevant though, and may need a replacement with site-specific ones
    • (?) Request analyzer (e.g., prediction of efficiency and number of requests)


  • 0.0.7 more Wikipedia API support (page metadata, page lists etc.)
  • 0.1.0 efficiency & robustness of existing features
  • 0.2.0 documentation and principal portability to other languages
  • (maybe) online client-side demo, using Pyodide?..


Why not use Wikidata (or other structured Wikipedia-based data source, like DBPedia)?

Wikidata is a massive effort to represent Wikipedia in a computable form. But currently, it contains much less data than Wikipedia itself; and much less accessible for investigatory data extraction (TODO: good examples!) While it gets improved constantly, I wanted to tackle the problem from a different angle and see how accessible we can make Wikipedia itself, with all of its semi-structuredness.

Why not fetch and parse page sources in Wikitext?

Some similar projects (say, wtf_wikipedia) work by fetching page source in Wikitext form, and parsing it for data extraction. This road looks pretty tempting (and for several years, I went it myself with the previous iteration: infoboxer Ruby project). The problem here is that at first sight, Wikitext is better structured: large chunks of data are represented by templates like {{Infobox; field=value, ...}} so it really seems like a better source for data extraction. There are two huge problems with this approach, though:

  1. The list of templates is infinite and unpredictable. While ten city-related pages would have {{Infobox city in a pretty similar form, the eleventh will have {{Geobox capital with all the different fields and conventions—but in HTML they would render to the similarly-looking <table class="infobox". Or, some TV series will represent a list of episodes with just a plain table markup, while the other will use a sophisticated {{Episode list template. And it all might change with time (some spring cleanup replacing all the template names or converting some regular text to a template). The HTML version is much more stable and predictable.
  2. To library/QL users, having Wikitext&templates as the base would mean that writing queries requires intimate knowledge of Wikitext format and conventions and frequently looking at the page's source (via "Edit") to find out how something is represented in Wikitext. OTOH, HTML/CSS-based approach relies on widely known elements like headers, lists, links, tables, and simple grouping objects, like sections and sentences. Most of them are unambiguously deduced by looking at the page in the browser or, in the worst case, by "inspect element" (to find out its particular class/id).

Prior work

This project is the N-th iteration of the ideas about providing "common knowledge" in a computable form. Most of the previous work was done in Ruby and centered around Reality project; and included, amongst other things, Infoboxer the Wikipedia parser/high-level client, and MediaWiktory the idiomatic low-level MediaWiki client. Some of that work still to be incorporated into WikipediaQL and sister projects.

That project was once inspired by "integrated knowledge" feature of Wolfram Language, I've talked about it (and other topics leading to this project) in a Twitter thread (yes).

The WikipediaQL syntax seems to be subconsciously inspired by qsx selectors language. (By subconsciously I mean I don't remember thinking "Oh, I should do something similar", but the day I've published WikipediQL, service have reminded me I starred qsx in December 2020. I started to think about WikipediaQL syntax in June 2021, but there are striking similarities, so it should be related to some indirect inspiration by that project.)



Open Source Agenda is not affiliated with "Wikipedia Ql" Project. README Source: zverok/wikipedia_ql
Open Issues
Last Commit
7 months ago

Open Source Agenda Badge

Open Source Agenda Rating