Monday, June 29, 2009

Wikipedia taxonomy, the good, the bad, and the very ugly

In the previous post I suggested that a productive way to meet EOL's goal of a web page per taxon would be to build upon Wikipedia, rather than go it alone. In a nutshell the arguments were:

  1. Wikipedia has considerable traction and has some richly populated taxon pages

  2. The linked data community uses DBPedia.org as a core source of URIs for entities, as as DBPedia is derived from Wikipedia the later will be the core source of identifiers for taxa

To explore this a little further I grabbed two files from the 20090618 Wikipedia dump, namely page.sql and templatelinks.sql, and extracted page ids and titles for Wikipedia pages containing the Taxobox template. I then queried Wikipedia for the source for each of these pages, and tried to extract the taxonomic information from each page (a tedious and error-prone process at best).

I've put together a shockingly crude web page where you can browse the results (warning, this page is a 10 minute hack with little error checking).

There is some good news. There are over 120,000 taxon pages (I've not got an exact figure because the Taxobox template occurs on some pages that aren't taxon pages, such as documentation and user pages). Some pages are extensive (the largest page is Dinosaur for which the source text is 128K in size), and there are lots of links to external references (I counted 7205 distinct DOIs to papers and/or books, and 3248 distinct ISBNs). This represents a degree of external linkage that puts EOL to shame.

However, there are also some major problems. Firstly, Wikipedia does not have a single, internally consistent classification (i.e., the classification is not a tree). This is not unexpected, given that Wikipedia pages comprise semi-structured text that is (largely) manually entered. It's not a database. If it were, the simplest way to ensure consistency would be to have each child node include a pointer to its parent, and when we want a list of the children of the parent node we simply query the database ("what nodes have this node as their parent?"). Because Wikipedia isn't a database, authors have entered these two relationships ("has parent" and "has child") on different pages, and these often conflict. For a spectacular example of this, take a look at the page for Amphibia. When I scrapped Wikipedia I extracted the "has parent" link, as this is the simplest way to create a tree. This results in over 200 child taxa for Amphibia, yet the Wikipedia page for Amphibia lists only four child taxa. What appears to be happening is that many fossil taxa are being added to Wikipedia, and since we are often hazy about where they go in the tree, authors are listing their parent taxon as (in this case) "Amphibia". Given this direct link, they should also be listed as children of Amphibia (although, of course, that would make a mess of the Amphibia page). Perhaps the solution is to add a "incerta sedis" taxon page for each taxon, and make that the parent of all the taxa that we're aren't sure where to put. This would ensure consistency, but not make the current taxon pages unreadable.

Homonymy (the same name for different taxa) also raises it's ugly head. For example, the page for the crab family Latreilliidae lists the genus Latreillia, which is a fly. In this case, the fly genus Latreillia Robineau-Desvoidy, is a junior homonym of the crab genus Latreillia Roux (see http://biodiversitylibrary.org/page/12221111).

Finally, the page titles (which become the basis of DBPedia.org URIs) are a muddled mixture of common and scientific names.

So, what to do? Well, the idea of simply using Wikipedia as is isn't going to fly, it's too broken. We will have to contemplate a concerted effort to fix it (which will require using bots to clean up the inconsistencies). Another option (assuming that we like the Wiki-style environment) is to use a semantic wiki (see my earlier post), which constrains some of the possible markup, but retains a lot of the freedom that make wikis so powerful.

This isn't an argument for not using Wikipedia as such, it's arguably still much more informative than, say, EOL. It's just that it's showing signs of the limitations of free-form text entry. The trick is to find a way to combine the obvious strengths of this approach (ease of creating and editing pages, massive community support) with the more structured approach needed to avoid the internal inconsistencies that currently bedevil Wikipedia.