I agree that the existing rating of articles are not very useful. Many articles are unassessed. Others were assessed within days of the article being created and assessed as Stub/Start and have not been revisited since despite considerable further development of the article. Some people work very hard to get an article to GA (or whatever) and explicitly request assessment. I would think most "high quality" articles have had people actively working to achieve a high rating and explicitly requesting assessment. I don't know how many articles get to high levels of quality just through the uncoordinated contributions of the crowd but I bet it's hardly enough. Indeed, I suspect if you train on high quality articles, you'll learn that having a small number of editors doing a lot of work in its recent history is the best indicator of quality.
If you are going to train your heuristics, I'd suggest collecting articles which have had little/no further development since their last rating so that you know the assessments have some chance of being accurate.
I doubt there is any single metric that is a predictor of quality but I think citations is probably a good proxy. Of course, there are probably counter-examples but generally an article with lots of citations suggests a sincere effort at a better-quality article. Of course if any tool is deployed to automatically assess article quality, then we can expect people to "game" it, but at this stage one would assume that people are not actively gaming the rating system while it has a manual assessment process. However, people probably are "gaming" NPOV in specific articles by adding lots of citations that support their views; I doubt any metric will allow you to easily spot this kind of behaviour without doing some kind of analysis of the sources and interrelationships between them.
But, as Laura comments, there may be a lot of citations clustered in a small part of the article, but few elsewhere. Also, the number of sources is relevant - I can cite the same source 1000 times in one article and that's probably not quality either. I'd be inclined to reduce the influence of both multiple citations at the same point of the text (or very close in the text) as well as repeated citations to the same source. It's not that either is bad but there should be some limit to how much they influence any conclusions.
Kerry
_____
From: wiki-research-l-bounces@lists.wikimedia.org [mailto:wiki-research-l-bounces@lists.wikimedia.org] On Behalf Of WereSpielChequers Sent: Sunday, 15 December 2013 6:54 PM To: Research into Wikimedia content and communities Subject: Re: [Wiki-research-l] Existitng Research on Article QualityHeuristics?
Re other dimensions or heuristics:
Very few articles are rated as Featured, and not that many as Good, if you are going to use https://en.wikipedia.org/wiki/Wikipedia:Version_1.0_Editorial_Team/Assessme nt that rating system I'd suggest also including the lower levels, and indeed whether an article has been assessed and typically how long it takes for a new article to be assessed. Uganda for example has 1 https://en.wikipedia.org/wiki/Wikipedia:UGANDA#Recognized_content Featured article, 3 Good Articles and nearly 400 unassessed on the English language Wikipedia.
For a crowd sourced project like Wikipedia the size of the crowd is crucial and varies hugely per article. So I'd suggest counting the number of different editors other than bots who have contributed to the article. It might also be worth getting some measure of local internet speed or usage level as context. There was a big upgrade to East Africa's Internet connection a few years ago. For Wikipedia the crucial metric is the size of the Internet comfortable population with some free time and ready access to PCs, I'm not sure we've yet measured how long it takes from people getting internet access to their being sufficiently confident to edit Wikipedia articles, I suspect the answer is age related, but it would be worth checking the various editor surveys to see if this has been collected yet. My understanding is that in much of Africa many people are bypassing the whole PC thing and going straight to smartphones, and of course for mobilephone users Wikipedia is essentially a queryable media rather than an interactive editable one.
Whether or not a Wikipedia article has references is a quality dimension you might want to look at. At least on EN it is widely assumed to be a measure of quality, though I don't recall ever seeing a study of the relative accuracy of cited and uncited Wikipedia information.
Thankfully the Article Feedback tool has been almost eradicated from the English language Wikipedia, I don't know if it is still on French or Swahili. I don't see it as being connected to the quality of article, thouugh it should be an interesting measure of how loved or hated a given celebrity was during the time the tool was deployed. So I'd suggest ignoring it in your research on article quality.
Hope that helps
Jonathan
On 15 December 2013 06:15, Klein,Max kleinm@oclc.org wrote:
Wiki Research Junkies,
I am investigating the comparative quality of articles about Cote d'Ivoire and Uganda versus other countries. I wanted to answer the question of what makes high-quality articles? Can anyone point me to any existing research on heuristics of Article Quality? That is, determining an articles quality by the wikitext properties, without human rating? I would also consider using data from the Article Feedback Tools, if there were dumps available for each Article in English, French, and Swahili Wikipedias. This is all the raw data I can seem to find http://toolserver.org/~dartar/aft5/dumps/
The heuristic technique that I currently using is training a naive Bayesian filter based on:
* Per Section.
o Text length in each section
o Infoboxes in each section.
* Filled parameters in each infobox
o Images in each section
* Good Article, Featured Article?
* Then Normalize on Page Views per on population / speakers of native language
Can you also think of any other dimensions or heuristics to programatically rate?
Best,
Maximilian Klein Wikipedian in Residence, OCLC +17074787023 tel:%2B17074787023
_______________________________________________ Wiki-research-l mailing list Wiki-research-l@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/wiki-research-l