So here's the list of accounts that were used
in order to create the articles:
Also some edits may have been done through IPs.
In discussion with Sidd it was clear that they did not plan to ever mass-create a large
number of articles, and it is only these 50 articles or so we can clean up now. I am not
terribly worried about this particular work (according to the paper there were 47
surviving articles at the time of writing, i.e. in Spring).
What I am concerned about is the fact that there will be more such experiments from other
groups. It would be great to set up a few rules for this kind of behavior, so that we can
at least point to them. If the only rule that was broken here was the "don't use
multiple accounts" rule, I am not sure whether that would be sufficient.
Cheers,
Denny
On Wed, Aug 10, 2016 at 1:47 AM Stuart A. Yeates
<syeates(a)gmail.com> wrote:
* The previous work you cite appears to have created articles in the draft namespace
rather than the article namespace. This is a very important and very relevant detail,
meaning your situation is in no way comparable to the previous work from my point of view
* You appear to be solving a problem that the community of wikipedia editors does not
have. We have enough low-quality stub articles that need human effort to improve and
we're not really interested in more unless either (a) they demonstrably combat some of
the systematic biases we're struggling with or (b) they demonstrably attract new
cohorts users to do that improvement. Note that the examples discussed in the research
newsletter are a non-English writer and a women writer. These are important details.
* Your paper appears not to attempt to make any attempt to measure the statistical
significance of your results; this isn't science.
* Most of your sources are _really_ _really_ bad.
https://en.wikipedia.org/wiki/Talonid
Contains 8 unique refs, one of which is good, one of which is a passable and the others
should be removed immediately (but I won't because it'll make it harder for third
parties reading this conversation to follow it.).
If you want to properly evaluate your technique, try this: Randomly pick N articles from
https://en.wikipedia.org/wiki/Category:Articles_lacking_sources subcats splitting them
into control and subjects randomly. Parse each subject article for sentences that your
system appears to understand. For each sentence your thing you understand look for
reliable sources to support that sentence. Add a single ref to a single statement in each
article. Add all the refs using a single account with a message on the user page about the
nature of the edits. If you're not able to add any refs, mark it as a failure. Measure
article lifespan for each group.
If you're in a hurry and want fast results, work with articles less than a week old
(hint: articles IDs are numerically increasing sequence) or the intersection of
https://en.wikipedia.org/wiki/Category:Articles_lacking_sources subcats and
Category:Articles_for_deletion Both of these groups of articles are actively being
considered for deletion.
cheers
stuart
--
...let us be heard from red core to black sky
> On Wed, Aug 10, 2016 at 9:30 AM, siddhartha banerjee <sidd2006(a)gmail.com>
wrote:
> Hello Everyone,
>
> I am the first author of the paper that Denny has referred. Firstly, I want to thank
Denny for asking me to join this list and know more about this discussion.
>
> 1. Regarding quality, we know that there are issues, and even in the conference, I
have repeatedly told the audience that I am not satisfied with the quality of the content
generated. However, the percentage of articles that were not removed when the paper was
submitted was minimal. I have sent Denny a list of accounts that were used and it might
have been possible that several articles created have been removed from those accounts
within the last couple of months. I was not aware of the multiple account policy.
>
> 2. The area of Wikipedia article generation have been explored by others in the past.
[
http://www.aclweb.org/anthology/P09-1024,
http://wwwconference.org/proceedings/www2011/companion/p161.pdf] We were not aware of any
rules regarding these sort of experiments. However, we do understand that such experiments
can harm the general quality of this great encyclopedic resource, hence we did out
analysis on bare minimum articles. In fact, we did our initial work on it back in 2014,
and Wikimedia research even covered details about our paper here --
https://blog.wikimedia.org/2015/02/02/wikimedia-research-newsletter-january…
>
> If questions were raised at that point, we would surely not have done anything
further on this, or rather do things offline without creating or adding any content on
Wikipedia.
>
> I understand your point about imposing rules and I think it makes sense. However,
during this research, we were not aware of any rules, hence continued our work.
> As I have told Denny, our purpose was to check whether we could create bare minimal
articles which could be eventually improved by authors on Wikipedia, and also to see if
they are totally removed. But, it was done with a few articles and we did not create
anything beyond that point. Also, we did not do any manual modifications to the articles
although we saw quality issues because it would void our analysis and claims.
>
> Thanks everyone for your time and the great work you are doing for the Wikipedia
community.
>
> Regards,
> Sidd
>
>
>
>
>
> _______________________________________________
> Wiki-research-l mailing list
> Wiki-research-l(a)lists.wikimedia.org
>
https://lists.wikimedia.org/mailman/listinfo/wiki-research-l
_______________________________________________
Wiki-research-l mailing list
Wiki-research-l(a)lists.wikimedia.org
https://lists.wikimedia.org/mailman/listinfo/wiki-research-l
_______________________________________________
Wiki-research-l mailing list
Wiki-research-l(a)lists.wikimedia.org