Today we are announcing https://blog.wikimedia.org/2017/02/07/scaling-understanding-of-harassment/ the first results of the collaboration between Wikimedia Research and Jigsaw on modeling personal attacks and other forms of harassment on English Wikipedia. We have released https://figshare.com/projects/Wikipedia_Talk/16731 a corpus of 95M user and article talk page comments as well as over 1M human labels produced by 4000 crowd-workers for a set of 100k comments. Documentation on our methodology and future work can be found in our paper Ex Machina: Personal Attacks Seen at Scale https://arxiv.org/abs/1610.08914 (to appear at WWW2017) and on our project page on meta https://meta.wikimedia.org/wiki/Research:Detox. If you are interested in contributing to the project, please get in touch via the project talk page https://meta.wikimedia.org/wiki/Research_talk:Detox. Another great way to get involved is to label a set of comment in the Wikilabels discussion quality campaign http://labels.wmflabs.org/ui/enwiki/.
I'm reflecting on this work and how awesome it was. I see that it's continued in our annual plan under the Community Health Initiative, but I am afraid it's taking a secondary role without Ellery and others to drive it. On https://meta.wikimedia.org/wiki/Community_health_initiative/AbuseFilter it's only featured as a question under the #Functionality section.
I just wanted to point this out and offer to help if I can be of use.
On Tue, Feb 7, 2017 at 5:16 PM, Ellery Wulczyn ewulczyn@wikimedia.org wrote:
Today we are announcing https://blog.wikimedia.org/2017/02/07/scaling-understanding-of-harassment/ the first results of the collaboration between Wikimedia Research and Jigsaw on modeling personal attacks and other forms of harassment on English Wikipedia. We have released https://figshare.com/projects/Wikipedia_Talk/16731 a corpus of 95M user and article talk page comments as well as over 1M human labels produced by 4000 crowd-workers for a set of 100k comments. Documentation on our methodology and future work can be found in our paper Ex Machina: Personal Attacks Seen at Scale https://arxiv.org/abs/1610.08914 (to appear at WWW2017) and on our project page on meta https://meta.wikimedia.org/wiki/Research:Detox. If you are interested in contributing to the project, please get in touch via the project talk page https://meta.wikimedia.org/wiki/Research_talk:Detox. Another great way to get involved is to label a set of comment in the Wikilabels discussion quality campaign http://labels.wmflabs.org/ui/enwiki/.
Analytics mailing list Analytics@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/analytics
Hi Dan -- we are actually in touch with Detox as part of the Community Health initiative. They are doing their first quarterly check in this quarter so expect some updates then. Ping me offlist if you want more info.
-Toby
On Wed, Jun 21, 2017 at 10:48 AM, Dan Andreescu dandreescu@wikimedia.org wrote:
I'm reflecting on this work and how awesome it was. I see that it's continued in our annual plan under the Community Health Initiative, but I am afraid it's taking a secondary role without Ellery and others to drive it. On https://meta.wikimedia.org/wiki/Community_health_initiative/AbuseFilter it's only featured as a question under the #Functionality section.
I just wanted to point this out and offer to help if I can be of use.
On Tue, Feb 7, 2017 at 5:16 PM, Ellery Wulczyn ewulczyn@wikimedia.org wrote:
Today we are announcing <https://blog.wikimedia.org/2017/02/07/scaling-
understanding-of-harassment/> the
first results of the collaboration between Wikimedia Research and Jigsaw
on
modeling personal attacks and other forms of harassment on English Wikipedia. We have released https://figshare.com/projects/Wikipedia_Talk/16731 a corpus of 95M
user
and article talk page comments as well as over 1M human labels produced
by
4000 crowd-workers for a set of 100k comments. Documentation on our methodology and future work can be found in our paper Ex Machina: Personal Attacks Seen at Scale https://arxiv.org/abs/1610.08914 (to appear at WWW2017) and on our project page on meta https://meta.wikimedia.org/wiki/Research:Detox. If you are interested in contributing to the project, please get in touch via the project talk page https://meta.wikimedia.org/wiki/Research_talk:Detox. Another
great
way to get involved is to label a set of comment in the Wikilabels discussion quality campaign http://labels.wmflabs.org/ui/enwiki/.
Analytics mailing list Analytics@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/analytics
Wiki-research-l mailing list Wiki-research-l@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/wiki-research-l
Hi Dan,
Thanks for your note. :)
On the Research end, Dario is still a big supporter of the efforts around research to help us better understand harassment (as you noticed in our commitments to the annual plan) and with Ellery's departure, I've been helping him a bit to make sure we can move forward on this front. More specifically, and while we're continuing the research with Nithum and Lucas who were Ellery's collaborators on the Detox project, we recently initiated https://meta.wikimedia.org/wiki/Research:Study_of_harassment_and_its_impact with Cristian and Yiqing from Cornell University. We are very excited about this new collaboration as Cristian has years of experience in spaces that are very relevant to the socio-technical problems related to harassment. I think you will enjoy reading that page which signal the early directions of the research.
The whole harassment research team meets every 2 weeks, if you're curious what's going on on this front and on our end and you want to listen in, please ping me. And, thank you for the offer to help. We may take you up on that. :)
Best, Leila
-- Leila Zia Senior Research Scientist Wikimedia Foundation
On Wed, Jun 21, 2017 at 7:55 PM, Toby Negrin tnegrin@wikimedia.org wrote:
Hi Dan -- we are actually in touch with Detox as part of the Community Health initiative. They are doing their first quarterly check in this quarter so expect some updates then. Ping me offlist if you want more info.
-Toby
On Wed, Jun 21, 2017 at 10:48 AM, Dan Andreescu dandreescu@wikimedia.org wrote:
I'm reflecting on this work and how awesome it was. I see that it's continued in our annual plan under the Community Health Initiative, but I am afraid it's taking a secondary role without Ellery and others to drive it. On https://meta.wikimedia.org/wiki/Community_health_initiative/AbuseFilter it's only featured as a question under the #Functionality section.
I just wanted to point this out and offer to help if I can be of use.
On Tue, Feb 7, 2017 at 5:16 PM, Ellery Wulczyn ewulczyn@wikimedia.org wrote:
Today we are announcing
https://blog.wikimedia.org/2017/02/07/scaling-understanding-of-harassment/ the first results of the collaboration between Wikimedia Research and Jigsaw on modeling personal attacks and other forms of harassment on English Wikipedia. We have released https://figshare.com/projects/Wikipedia_Talk/16731 a corpus of 95M user and article talk page comments as well as over 1M human labels produced by 4000 crowd-workers for a set of 100k comments. Documentation on our methodology and future work can be found in our paper Ex Machina: Personal Attacks Seen at Scale https://arxiv.org/abs/1610.08914 (to appear at WWW2017) and on our project page on meta https://meta.wikimedia.org/wiki/Research:Detox. If you are interested in contributing to the project, please get in touch via the project talk page https://meta.wikimedia.org/wiki/Research_talk:Detox. Another great way to get involved is to label a set of comment in the Wikilabels discussion quality campaign http://labels.wmflabs.org/ui/enwiki/.
Analytics mailing list Analytics@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/analytics
Wiki-research-l mailing list Wiki-research-l@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/wiki-research-l
Analytics mailing list Analytics@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/analytics
I'm glad that work on detecting and addressing harassment are moving forward.
At the same time, I'd appreciate getting a more precise understanding of how WMF is defining the word "harassment". There are legal definitions and dictionary definitions, but I don't think that there is One Definition to Rule Them All. I'm hoping that WMF will be careful to distinguish debate and freedom to express opinions from harassment; we may disagree with minority or fringe views (even views that are offensive to some) but that doesn't necessarily mean that we should use policy and admin tools instead of persuasion and other tools (such as content policies about verifiability and notability) to address them (and in some cases Wikipedia may not be a good place for these discussions). Other distinctions include (1) the distinction between a personal attack and harassment ( https://blog.wikimedia.org/2017/02/07/scaling-understanding-of-harassment/ appears to have equivocated the two definitions, while English Wikipedia policy makes distinctions between them), and (2) the distinction between a personal attack and an evidence-based critique.
Also note that definitions of what constitutes an attack may vary between languages; for example an expression which sounds insulting to someone in one place, culture, or language may mean something very different or relatively benign in a different place, culture, or language. I had an experience myself when I made a statement to someone which from my perspective was a statement of fact, and the other party took it as an insult. I don't apologize for what I said since from my perspective it was valid, and the other party has not apologized for their reaction, but the point is that defining what constitutes a personal attack or harassment can be a very subjective business and I'm not sure to what extent I would trust an AI to evaluate what constitutes a personal attack or harassment in a wide range of contexts. I get the impression that WMF intends to flag potentially problematic edits for admins to review, which I think could be a good thing, but I hope that there is great care being invested in how the AI is being trained to define personal attacks and harassment, and I wouldn't necessarily want admins to be encouraged to substitute the opinion of an AI for their own.
I understand the desire to tone down some of the more heated discourse around Wikipedia for the sake of improving our user population statistics, and at the same time I'm hoping that we can continue to have very strong support for freedom of expression and differences of opinion. This is a difficult balancing act. I think that moving the needle a bit in the direction of more civility would be a good thing, but I get the impression that there are plenty of edits that are blatant personal attacks that we don't need to move the needle a lot, and could instead focus on more rapidly and thoroughly addressing instances where there is ample evidence that people's intentions were malicious.
Pine
On Wed, Jun 21, 2017 at 2:08 PM, Leila Zia leila@wikimedia.org wrote:
Hi Dan,
Thanks for your note. :)
On the Research end, Dario is still a big supporter of the efforts around research to help us better understand harassment (as you noticed in our commitments to the annual plan) and with Ellery's departure, I've been helping him a bit to make sure we can move forward on this front. More specifically, and while we're continuing the research with Nithum and Lucas who were Ellery's collaborators on the Detox project, we recently initiated https://meta.wikimedia.org/wiki/Research:Study_of_ harassment_and_its_impact with Cristian and Yiqing from Cornell University. We are very excited about this new collaboration as Cristian has years of experience in spaces that are very relevant to the socio-technical problems related to harassment. I think you will enjoy reading that page which signal the early directions of the research.
The whole harassment research team meets every 2 weeks, if you're curious what's going on on this front and on our end and you want to listen in, please ping me. And, thank you for the offer to help. We may take you up on that. :)
Best, Leila
-- Leila Zia Senior Research Scientist Wikimedia Foundation
On Wed, Jun 21, 2017 at 7:55 PM, Toby Negrin tnegrin@wikimedia.org wrote:
Hi Dan -- we are actually in touch with Detox as part of the Community Health initiative. They are doing their first quarterly check in this quarter so expect some updates then. Ping me offlist if you want more
info.
-Toby
On Wed, Jun 21, 2017 at 10:48 AM, Dan Andreescu <
dandreescu@wikimedia.org>
wrote:
I'm reflecting on this work and how awesome it was. I see that it's continued in our annual plan under the Community Health Initiative, but
I
am afraid it's taking a secondary role without Ellery and others to
drive
it. On https://meta.wikimedia.org/wiki/Community_health_initiative/AbuseFilter it's only featured as a question under the #Functionality section.
I just wanted to point this out and offer to help if I can be of use.
On Tue, Feb 7, 2017 at 5:16 PM, Ellery Wulczyn ewulczyn@wikimedia.org wrote:
Today we are announcing
understanding-of-harassment/>
the first results of the collaboration between Wikimedia Research and
Jigsaw
on modeling personal attacks and other forms of harassment on English Wikipedia. We have released https://figshare.com/projects/Wikipedia_Talk/16731 a corpus of 95M user and article talk page comments as well as over 1M human labels
produced
by 4000 crowd-workers for a set of 100k comments. Documentation on our methodology and future work can be found in our paper Ex Machina: Personal Attacks Seen at Scale https://arxiv.org/abs/1610.08914 (to appear at WWW2017) and on our project page on meta https://meta.wikimedia.org/wiki/Research:Detox. If you are
interested
in contributing to the project, please get in touch via the project
talk
page https://meta.wikimedia.org/wiki/Research_talk:Detox. Another great way to get involved is to label a set of comment in the Wikilabels discussion quality campaign http://labels.wmflabs.org/ui/enwiki/.
Analytics mailing list Analytics@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/analytics
Wiki-research-l mailing list Wiki-research-l@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/wiki-research-l
Analytics mailing list Analytics@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/analytics
Analytics mailing list Analytics@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/analytics
Hi Pine,
On Thu, Jun 22, 2017 at 2:03 AM, Pine W wiki.pine@gmail.com wrote:
At the same time, I'd appreciate getting a more precise understanding of how WMF is defining the word "harassment".
This is a policy question and wiki-research-l and analytics mailing lists are not the best place to discuss it. (I'm not sure where it is, maybe policy public list?)
If you are interested about learning how a specific research effort has handled this question, I suggest you reach out to the researchers in that effort. In the case of Ex Machina: Personal Attacks Seen at Scale https://arxiv.org/abs/1610.08914, Section 3 should give you a relatively detailed description of how this question was approached.
Best, Leila
Pine
On Wed, Jun 21, 2017 at 2:08 PM, Leila Zia leila@wikimedia.org wrote:
Hi Dan,
Thanks for your note. :)
On the Research end, Dario is still a big supporter of the efforts around research to help us better understand harassment (as you noticed in our commitments to the annual plan) and with Ellery's departure, I've been helping him a bit to make sure we can move forward on this front. More specifically, and while we're continuing the research with Nithum and Lucas who were Ellery's collaborators on the Detox project, we recently initiated https://meta.wikimedia.org/wiki/Research:Study_of_harassment _and_its_impact with Cristian and Yiqing from Cornell University. We are very excited about this new collaboration as Cristian has years of experience in spaces that are very relevant to the socio-technical problems related to harassment. I think you will enjoy reading that page which signal the early directions of the research.
The whole harassment research team meets every 2 weeks, if you're curious what's going on on this front and on our end and you want to listen in, please ping me. And, thank you for the offer to help. We may take you up on that. :)
Best, Leila
-- Leila Zia Senior Research Scientist Wikimedia Foundation
On Wed, Jun 21, 2017 at 7:55 PM, Toby Negrin tnegrin@wikimedia.org wrote:
Hi Dan -- we are actually in touch with Detox as part of the Community Health initiative. They are doing their first quarterly check in this quarter so expect some updates then. Ping me offlist if you want more
info.
-Toby
On Wed, Jun 21, 2017 at 10:48 AM, Dan Andreescu <
dandreescu@wikimedia.org>
wrote:
I'm reflecting on this work and how awesome it was. I see that it's continued in our annual plan under the Community Health Initiative,
but I
am afraid it's taking a secondary role without Ellery and others to
drive
it. On https://meta.wikimedia.org/wiki/Community_health_initiative/
AbuseFilter
it's only featured as a question under the #Functionality section.
I just wanted to point this out and offer to help if I can be of use.
On Tue, Feb 7, 2017 at 5:16 PM, Ellery Wulczyn <ewulczyn@wikimedia.org
wrote:
Today we are announcing
<https://blog.wikimedia.org/2017/02/07/scaling-understanding
-of-harassment/>
the first results of the collaboration between Wikimedia Research and
Jigsaw
on modeling personal attacks and other forms of harassment on English Wikipedia. We have released https://figshare.com/projects/Wikipedia_Talk/16731 a corpus of 95M user and article talk page comments as well as over 1M human labels
produced
by 4000 crowd-workers for a set of 100k comments. Documentation on our methodology and future work can be found in our paper Ex Machina: Personal Attacks Seen at Scale https://arxiv.org/abs/1610.08914
(to
appear at WWW2017) and on our project page on meta https://meta.wikimedia.org/wiki/Research:Detox. If you are
interested
in contributing to the project, please get in touch via the project
talk
page https://meta.wikimedia.org/wiki/Research_talk:Detox. Another great way to get involved is to label a set of comment in the Wikilabels discussion quality campaign http://labels.wmflabs.org/ui/enwiki/.
Analytics mailing list Analytics@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/analytics
Wiki-research-l mailing list Wiki-research-l@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/wiki-research-l
Analytics mailing list Analytics@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/analytics
Analytics mailing list Analytics@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/analytics
Analytics mailing list Analytics@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/analytics
I agree you can probably never pin down these terms to everyone's satisfaction. But, at the end of the day, is the real issue here the definition of harassment or is it the issue of people leaving Wikipedia because of unpleasant interactions with other people or perhaps retaliating in some inappropriate way. Harassment may not even be occurring on a Talk page. If someone stalks you on-wiki and reverts each of your edits, you are probably being harassed without a word being said on Talk.
This is the problem. Two people can see the same set of events or the same commentary from very different points of view. The question of "harassment" isn't completely decidable in the real world for the same reasons. But if we train the algorithms based on human assessments (provided that a wide range of people were making those assessments), we do have something useful to work with to begin to test hypotheses in the lab before taking real-world action.
For example, I find it very interesting that a small group of experienced users appear responsible for a lot of apparently obvious personal attacks. It does indeed suggest that these people think themselves unstoppable, whether that is being they believe themselves "unblockable" or perhaps they feel safe in the knowledge that their less-experienced victim is unlikely to know how to complain. Or perhaps they are just bantering among themselves, like a bunch of mate at the pub? But it certainly seems to suggest that there is a way to start identifying potential problem users for a human-based investigation.
But does the "community" really care about harassment to investigate them? Would it really take action against experienced users who engaged in harassment? Past events suggest not.
Kerry
-----Original Message----- From: Wiki-research-l [mailto:wiki-research-l-bounces@lists.wikimedia.org] On Behalf Of Pine W Sent: Thursday, 22 June 2017 10:04 AM To: A mailing list for the Analytics Team at WMF and everybody who has an interest in Wikipedia and analytics. analytics@lists.wikimedia.org; Wiki Research-l wiki-research-l@lists.wikimedia.org Subject: Re: [Wiki-research-l] [Analytics] Wikipedia Detox: Scaling up our understanding of harassment on Wikipedia
I'm glad that work on detecting and addressing harassment are moving forward.
At the same time, I'd appreciate getting a more precise understanding of how WMF is defining the word "harassment". There are legal definitions and dictionary definitions, but I don't think that there is One Definition to Rule Them All. I'm hoping that WMF will be careful to distinguish debate and freedom to express opinions from harassment; we may disagree with minority or fringe views (even views that are offensive to some) but that doesn't necessarily mean that we should use policy and admin tools instead of persuasion and other tools (such as content policies about verifiability and notability) to address them (and in some cases Wikipedia may not be a good place for these discussions). Other distinctions include (1) the distinction between a personal attack and harassment ( https://blog.wikimedia.org/2017/02/07/scaling-understanding-of-harassment/ appears to have equivocated the two definitions, while English Wikipedia policy makes distinctions between them), and (2) the distinction between a personal attack and an evidence-based critique.
Also note that definitions of what constitutes an attack may vary between languages; for example an expression which sounds insulting to someone in one place, culture, or language may mean something very different or relatively benign in a different place, culture, or language. I had an experience myself when I made a statement to someone which from my perspective was a statement of fact, and the other party took it as an insult. I don't apologize for what I said since from my perspective it was valid, and the other party has not apologized for their reaction, but the point is that defining what constitutes a personal attack or harassment can be a very subjective business and I'm not sure to what extent I would trust an AI to evaluate what constitutes a personal attack or harassment in a wide range of contexts. I get the impression that WMF intends to flag potentially problematic edits for admins to review, which I think could be a good thing, but I hope that there is great care being invested in how the AI is being trained to define personal attacks and harassment, and I wouldn't necessarily want admins to be encouraged to substitute the opinion of an AI for their own.
I understand the desire to tone down some of the more heated discourse around Wikipedia for the sake of improving our user population statistics, and at the same time I'm hoping that we can continue to have very strong support for freedom of expression and differences of opinion. This is a difficult balancing act. I think that moving the needle a bit in the direction of more civility would be a good thing, but I get the impression that there are plenty of edits that are blatant personal attacks that we don't need to move the needle a lot, and could instead focus on more rapidly and thoroughly addressing instances where there is ample evidence that people's intentions were malicious.
Pine
On Wed, Jun 21, 2017 at 2:08 PM, Leila Zia leila@wikimedia.org wrote:
Hi Dan,
Thanks for your note. :)
On the Research end, Dario is still a big supporter of the efforts around research to help us better understand harassment (as you noticed in our commitments to the annual plan) and with Ellery's departure, I've been helping him a bit to make sure we can move forward on this front. More specifically, and while we're continuing the research with Nithum and Lucas who were Ellery's collaborators on the Detox project, we recently initiated https://meta.wikimedia.org/wiki/Research:Study_of_ harassment_and_its_impact with Cristian and Yiqing from Cornell University. We are very excited about this new collaboration as Cristian has years of experience in spaces that are very relevant to the socio-technical problems related to harassment. I think you will enjoy reading that page which signal the early directions of the research.
The whole harassment research team meets every 2 weeks, if you're curious what's going on on this front and on our end and you want to listen in, please ping me. And, thank you for the offer to help. We may take you up on that. :)
Best, Leila
-- Leila Zia Senior Research Scientist Wikimedia Foundation
On Wed, Jun 21, 2017 at 7:55 PM, Toby Negrin tnegrin@wikimedia.org wrote:
Hi Dan -- we are actually in touch with Detox as part of the Community Health initiative. They are doing their first quarterly check in this quarter so expect some updates then. Ping me offlist if you want more
info.
-Toby
On Wed, Jun 21, 2017 at 10:48 AM, Dan Andreescu <
dandreescu@wikimedia.org>
wrote:
I'm reflecting on this work and how awesome it was. I see that it's continued in our annual plan under the Community Health Initiative, but
I
am afraid it's taking a secondary role without Ellery and others to
drive
it. On https://meta.wikimedia.org/wiki/Community_health_initiative/AbuseFi lter it's only featured as a question under the #Functionality section.
I just wanted to point this out and offer to help if I can be of use.
On Tue, Feb 7, 2017 at 5:16 PM, Ellery Wulczyn ewulczyn@wikimedia.org wrote:
Today we are announcing
understanding-of-harassment/>
the first results of the collaboration between Wikimedia Research and
Jigsaw
on modeling personal attacks and other forms of harassment on English Wikipedia. We have released https://figshare.com/projects/Wikipedia_Talk/16731 a corpus of 95M user and article talk page comments as well as over 1M human labels
produced
by 4000 crowd-workers for a set of 100k comments. Documentation on our methodology and future work can be found in our paper Ex Machina: Personal Attacks Seen at Scale https://arxiv.org/abs/1610.08914 (to appear at WWW2017) and on our project page on meta https://meta.wikimedia.org/wiki/Research:Detox. If you are
interested
in contributing to the project, please get in touch via the project
talk
page https://meta.wikimedia.org/wiki/Research_talk:Detox. Another great way to get involved is to label a set of comment in the Wikilabels discussion quality campaign http://labels.wmflabs.org/ui/enwiki/.
Analytics mailing list Analytics@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/analytics
Wiki-research-l mailing list Wiki-research-l@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/wiki-research-l
Analytics mailing list Analytics@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/analytics
Analytics mailing list Analytics@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/analytics
_______________________________________________ Wiki-research-l mailing list Wiki-research-l@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/wiki-research-l