Hello,
Just yesterday I had a long talk with a researcher about how to define and detect trolls on Wikipedia. E.g., whether "unintentional trolling" should be included or not.
In my opinion, it is not possible to detect by machine trollism, unkindness, harassment, mobbing etc., maybe with the exception of swear words. A lot of turntaking, deviation from the topic and other phenomena can be experienced by the participants as positive or as negative. You might need to ask them, and even then they might not be aware of a problem that works through in subtlety. Also, third persons not involved in the conversation can be effected negatively (look at ... page X... and you know why you don't like to contribute there).
Kind regards Ziko
2015-11-15 17:40 GMT+01:00 Katherine Casey fluffernutter.wiki@gmail.com:
I'd be happy to offer my admin/oversighter experience and knowledge to help you develop the labeling and such, Aaron! I just commented on Andreas's proposal on the Community Wishlist, but to summarize here: I see a lot of potential pitfalls in trying to handle/generalize this with machine learning, but I also see a lot of potential value, and I think it's something we should be investigating.
-Fluffernutter
On Sun, Nov 15, 2015 at 11:32 AM, Aaron Halfaker ahalfaker@wikimedia.org wrote:
The League of Legends team collaborated with outside scientists to analyse their dataset. I would love to see the Wikimedia Foundation
engage
in a similar research project.
Oh! We are! :) When we have time. :\ One of the projects that I'd like to see done, but I've struggled to find the time for is a common talk page parser[1] that could produce a dataset of talk page interactions. I'd like this dataset to be easy to join to editor outcome measures. E.g. there might be "aggressive" talk that we don't know is problematic until we see the kind of effect that it has on other conversation participants.
Anyway, I want some powerful utilities and datasets out there to help academics look into this problem more easily. For revscoring, I'd like to be able to take a set of talk page diffs, have them classified in Wiki labels[2] as "aggressive" and the build a model for ORES[3] to be used however people see fit. You could then use ORES to do offline analysis of discussions for research. You could use ORES to interrupt the a user before saving a change. I'm sure there are other clever ideas that people have for what to do with such a model that I'm happy to enable it via the service. The hard part is getting a good dataset labeled.
If someone wants to invest some time and energy into this, I'm happy to work with you. We'll need more than programming help. We'll need a lot of help to figure out what dimensions we'll label talk page postings by and to do the actual labeling.
- https://github.com/Ironholds/talk-parser
- https://meta.wikimedia.org/wiki/Wiki_labels
- https://meta.wikimedia.org/wiki/ORES
On Sun, Nov 15, 2015 at 6:56 AM, Andreas Kolbe jayen466@gmail.com wrote:
On Sat, Nov 14, 2015 at 9:13 PM, Benjamin Lees emufarmers@gmail.com wrote:
This article highlights the happier side of things, but it appears that Lin's approach also involved completely removing bad actors: "Some players have also asked why we've taken such an aggressive stance when we've been focused on reform; well, the key here is that for most players, reform approaches are quite effective. But, for a number of players, reform attempts have been very unsuccessful which forces us to remove some of these players from League entirely."[0]
Thanks for the added context, Benjamin. Of course, banning bad actors
that
they consider unreformable is something Wikipedia admins have always done as well.
The League of Legends team began by building a dataset of interactions
that
the community considered unacceptable, and then applied machine-learning
to
that dataset.
It occurs to me that the English Wikipedia has ready access to such a dataset: it's the totality of revision-deleted and oversighted talk page posts. The League of Legends team collaborated with outside scientists to analyse their dataset. I would love to see the Wikimedia Foundation
engage
in a similar research project.
I've added this point to the community wishlist survey:
https://meta.wikimedia.org/wiki/2015_Community_Wishlist_Survey#Machine-learn...
P.S. As Rupert noted, over 90% of LoL players are male (how much over 90%?).[1] It would be interesting to know whether this percentage has changed along with the improvements described in the article.
Indeed. _______________________________________________ Wikimedia-l mailing list, guidelines at: https://meta.wikimedia.org/wiki/Mailing_lists/Guidelines Wikimedia-l@lists.wikimedia.org Unsubscribe: https://lists.wikimedia.org/mailman/listinfo/wikimedia-l, mailto:wikimedia-l-request@lists.wikimedia.org?subject=unsubscribe
Wikimedia-l mailing list, guidelines at: https://meta.wikimedia.org/wiki/Mailing_lists/Guidelines Wikimedia-l@lists.wikimedia.org Unsubscribe: https://lists.wikimedia.org/mailman/listinfo/wikimedia-l, mailto:wikimedia-l-request@lists.wikimedia.org?subject=unsubscribe
-- Karen Brown user:Fluffernutter
*Unless otherwise specified, any email sent from this address is in my volunteer capacity and does not represent the views or wishes of the Wikimedia Foundation* _______________________________________________ Wikimedia-l mailing list, guidelines at: https://meta.wikimedia.org/wiki/Mailing_lists/Guidelines Wikimedia-l@lists.wikimedia.org Unsubscribe: https://lists.wikimedia.org/mailman/listinfo/wikimedia-l, mailto:wikimedia-l-request@lists.wikimedia.org?subject=unsubscribe