User name policy states that "*bot*" names are reserved for bots. Thus, such a regex shouldn't be too hacky, but I cannot comment whether some non-automated cases might slip through new user patrol. I do think dumps make the 'users' table available, and I know for sure one could get a full list via the API. As a check on this, you could check that when these usernames edit, whether or not they set the "bot" flag. -AW

Very helpful, Lukas, I didn't know about the logging table. In some recent work [1] I found many users that appeared to be bots but whose edits did not have the bot flag set. My approach was to exclude users who didn't have a break of more than 6 hours between edits over the entire month I was studying. I was interested in the users who had multiple edit sessions in the month and so when with a straight threshold. A way to keep users with only one editing session would be to exclude users who have no break longer than X hours in an edit session lasting at least Y hours (e.g., a user who doesn't break for more than 6 hours in 5-6 days is probably not human) Cheers, Scott [1] Multilinguals and Wikipedia Editing http://www.scotthale.net/pubs/?websci2014 -- Scott Hale Oxford Internet Institute University of Oxford http://www.scotthale.net/ scott.hale(a)oii.ox.ac.uk On Sun, May 18, 2014 at 5:45 PM, Lukas Benedix <lbenedix(a)l3q.de> wrote: -- Scott Hale Oxford Internet Institute University of Oxford http://www.scotthale.net/ scott.hale(a)oii.ox.ac.uk _______________________________________________ Wiki-research-l mailing list Wiki-research-l(a)lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/wiki-research-l

Tsk tsk tsk, Brian. When the revolution comes, bot discriminators will get no mercy. :-) But seriously, my tl;dr: instead of asking if an account is or isn't a bot, ask if a set of edits are or are not automated Great responses so far: searching usernames for *bot will exclude non-bot users who were registered before the username policy change (although *Bot is a bit better), and the logging table is a great way to collect bot flags. However, Scott is right -- the bot flag (or *Bot username) doesn't signify a bot, it signifies a bureaucrat recognizing that a user account successfully went through the Bot Approval Group process. If I see an account with a bot flag, I can generally assume the edits that account makes are initiated by an automated software agent. This is especially the case in the main namespace. The inverse assumption is not nearly as easy: I can't assume that every edit made from an account *without* a bot flag was *not* an automated edit. About unauthorized bots: yes, there are a relatively small number of Wikipedians who, on occasion, run fully-automated, continuously-operating bots without approval. Complicating this, if someone is going to take the time to build and run a bot, but isn't going to create a separate account for it, then it is likely that they are also using that account to do non-automated edits. Sometimes new bot developers will run an unauthorized bot under their own account during the initial stages of development, and only later in the process will they create a separate bot account and seek formal approval and flagging. It can get tricky when you exclude all the edits from an account for being automated based on a single suspicious set of edits. More commonly, there are many more people who use automated batch tools like AutoWikiBrowser to support one-off tasks, like mass find-and-replace or category cleanup. Accounts powered by AWB are technically not bots, only because a human has to sit there and click "save" for every batch edit that is made. Some people will create a separate bot account for AWB work and get it approved and flagged, but many more will not bother. Then there are people using semi-automated, human-in-the-loop tools like Huggle to do vandal fighting. I find that the really hard question is whether you include or exclude these different kinds of 'cyborgs', because it really makes you think hard about what exactly you're measuring. Is someone who does a mass find-and-replace on all articles in a category a co-author of each article they edit? Is a vandal fighter patrolling the recent changes feed with Huggle a co-author of all the articles they edit when they revert vandalism and then move on to the next diff? What about somebody using rollback in the web browser? If so, what is it that makes these entities authors and ClueBot NG not an author? When you think about it, user accounts are actually pretty remarkable in that they allow such a diverse set of uses and agents to be attributed to a single entity. So when it comes to identifying automation, I personally think it is better to shift the unit of analysis from the user account to the individual edit. A bot flag lets you assume all edits from an account are automated, but you can use a range of approaches to identifying sets of automated edits from non-flagged accounts. Then I have a set of regex SQL queries in the Query Library [1] which parses edit summaries for the traces that AWB, Huggle, Twinkle, rollback, etc. automatically leave by default. You can also use the edit session approach like Scott has suggested -- Aaron and I found a few unauthorized bots in our edit session study [2], and we were even using a more aggressive break, with no more than a 60 minute gap between edits. To catch short bursts of bulk edits, you could look at large numbers of edits made in a short period of time -- I'd say more than 7 main namespace edits a minute for 10 minutes would be a hard rate for even a very aggressive vandal fighter to maintain with Huggle. I'll conclude by saying that different kinds of automated editing techniques are different ways of participating in and contributing to Wikipedia. To systematically exclude automated edits is to remove a very important, meaningful, and heterogeneous kind of activity from view. These activities constitute a core part of what Wikipedia is, particularly those forms of automation which the community has explicitly authorized and recognized. Now, we researchers inevitably have to selectively reveal and occlude -- a co-authorship network based on main namespace edits also excludes talk page discussions and conflict resolution, and this also constitutes a core part of what Wikipedia is. It isn't wrong per se to exclude automated edits, and it is certainly much worse to not recognize that they exist at all. However, I always appreciate seeing how the analysis would be different if bots were not excluded. The fact that there are these weird users which absolutely dominate a co-authorship network graph if you don't filter them out is pretty amazing, at least to me. Best, Stuart [1] https://wiki.toolserver.org/view/MySQL_queries#Automated_tool_and_bot_edits [2] http://stuartgeiger.com/cscw13-labor-hours.pdf On Sun, May 18, 2014 at 10:08 AM, Scott Hale <computermacgyver(a)gmail.com>wrote;wrote: Very helpful, Lukas, I didn't know about the logging table. In some recent work [1] I found many users that appeared to be bots but whose edits did not have the bot flag set. My approach was to exclude users who didn't have a break of more than 6 hours between edits over the entire month I was studying. I was interested in the users who had multiple edit sessions in the month and so when with a straight threshold. A way to keep users with only one editing session would be to exclude users who have no break longer than X hours in an edit session lasting at least Y hours (e.g., a user who doesn't break for more than 6 hours in 5-6 days is probably not human) Cheers, Scott [1] Multilinguals and Wikipedia Editing http://www.scotthale.net/pubs/?websci2014 -- Scott Hale Oxford Internet Institute University of Oxford http://www.scotthale.net/ scott.hale(a)oii.ox.ac.uk On Sun, May 18, 2014 at 5:45 PM, Lukas Benedix <lbenedix(a)l3q.de> wrote: Here is a list of currently flagged bots: https://en.wikipedia.org/w/index.php?title=Special:ListUsers&offset=&am… Another good point to look for bots is here: https://en.wikipedia.org/w/index.php?title=Special%3APrefixIndex&prefix… You should also have a look at this pages to find former bots: https://en.wikipedia.org/wiki/Wikipedia:Bots/Status/inactive_bots_1 https://en.wikipedia.org/wiki/Wikipedia:Bots/Status/inactive_bots_2 And last but not least the logging table you can access via tool labs: SELECT DISTINCT(log_title) FROM logging WHERE log_action = 'rights' AND log_params LIKE '%bot%'; Lukas Am So 18.05.2014 18:34, schrieb Andrew G. West: _______________________________________________ Wiki-research-l mai

Tsk tsk tsk, Brian. When the revolution comes, bot discriminators will get no mercy. :-) But seriously, my tl;dr: instead of asking if an account is or isn't a bot, ask if a set of edits are or are not automated Great responses so far: searching usernames for *bot will exclude non-bot users who were registered before the username policy change (although *Bot is a bit better), and the logging table is a great way to collect bot flags. However, Scott is right -- the bot flag (or *Bot username) doesn't signify a bot, it signifies a bureaucrat recognizing that a user account successfully went through the Bot Approval Group process. If I see an account with a bot flag, I can generally assume the edits that account makes are initiated by an automated software agent. This is especially the case in the main namespace. The inverse assumption is not nearly as easy: I can't assume that every edit made from an account *without* a bot flag was *not* an automated edit. About unauthorized bots: yes, there are a relatively small number of Wikipedians who, on occasion, run fully-automated, continuously-operating bots without approval. Complicating this, if someone is going to take the time to build and run a bot, but isn't going to create a separate account for it, then it is likely that they are also using that account to do non-automated edits. Sometimes new bot developers will run an unauthorized bot under their own account during the initial stages of development, and only later in the process will they create a separate bot account and seek formal approval and flagging. It can get tricky when you exclude all the edits from an account for being automated based on a single suspicious set of edits. More commonly, there are many more people who use automated batch tools like AutoWikiBrowser to support one-off tasks, like mass find-and-replace or category cleanup. Accounts powered by AWB are technically not bots, only because a human has to sit there and click "save" for every batch edit that is made. Some people will create a separate bot account for AWB work and get it approved and flagged, but many more will not bother. Then there are people using semi-automated, human-in-the-loop tools like Huggle to do vandal fighting. I find that the really hard question is whether you include or exclude these different kinds of 'cyborgs', because it really makes you think hard about what exactly you're measuring. Is someone who does a mass find-and-replace on all articles in a category a co-author of each article they edit? Is a vandal fighter patrolling the recent changes feed with Huggle a co-author of all the articles they edit when they revert vandalism and then move on to the next diff? What about somebody using rollback in the web browser? If so, what is it that makes these entities authors and ClueBot NG not an author? When you think about it, user accounts are actually pretty remarkable in that they allow such a diverse set of uses and agents to be attributed to a single entity. So when it comes to identifying automation, I personally think it is better to shift the unit of analysis from the user account to the individual edit. A bot flag lets you assume all edits from an account are automated, but you can use a range of approaches to identifying sets of automated edits from non-flagged accounts. Then I have a set of regex SQL queries in the Query Library [1] which parses edit summaries for the traces that AWB, Huggle, Twinkle, rollback, etc. automatically leave by default. You can also use the edit session approach like Scott has suggested -- Aaron and I found a few unauthorized bots in our edit session study [2], and we were even using a more aggressive break, with no more than a 60 minute gap between edits. To catch short bursts of bulk edits, you could look at large numbers of edits made in a short period of time -- I'd say more than 7 main namespace edits a minute for 10 minutes would be a hard rate for even a very aggressive vandal fighter to maintain with Huggle. I'll conclude by saying that different kinds of automated editing techniques are different ways of participating in and contributing to Wikipedia. To systematically exclude automated edits is to remove a very important, meaningful, and heterogeneous kind of activity from view. These activities constitute a core part of what Wikipedia is, particularly those forms of automation which the community has explicitly authorized and recognized. Now, we researchers inevitably have to selectively reveal and occlude -- a co-authorship network based on main namespace edits also excludes talk page discussions and conflict resolution, and this also constitutes a core part of what Wikipedia is. It isn't wrong per se to exclude automated edits, and it is certainly much worse to not recognize that they exist at all. However, I always appreciate seeing how the analysis would be different if bots were not excluded. The fact that there are these weird users which absolutely dominate a co-authorship network graph if you don't filter them out is pretty amazing, at least to me. Best, Stuart [1] https://wiki.toolserver.org/view/MySQL_queries#Automated_tool_and_bot_edits [2] http://stuartgeiger.com/cscw13-labor-hours.pdf On Sun, May 18, 2014 at 10:08 AM, Scott Hale <computermacgyver(a)gmail.com>wrote;wrote: _______________________________________________ Wiki-research-l mailing list Wiki-research-l(a)lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/wiki-research-l

Another thought I had was that because many semi-automated tools such as Twinkle and AWB leave parenthetical annotations in their revision comments

Thanks for all the references and excellent advice so far! I've looked into the Hale Anti-Bot Method™, but because I've sampled my corpus on articles (based on category co-membership), the resulting groupby users gives these semi-automated users more "normal" distributions since their other contributions are censored. In other words, I see only a fraction of these users' contributions and thus the resulting time intervals I observe are spaced farther apart (more typical) than they actually are. It's not feasible for me to get 100k+ users' histories just for the purposes of cleaning up ~6k articles' histories. Another thought I had was that because many semi-automated tools such as Twinkle and AWB leave parenthetical annotations in their revision comments, would this be a relatively inexpensive way to filter out revisions rather than users? Some caveats, I'd like to get domain experts' feedback on. I'm not expecting settled research, just input from others' experiences munging the data. 1. Is the inclusion of this markup in revision comments optional? This is a concern that some users may enable or disable it, so I may end up biasing inclusion based on users' preferences. 2. How have these flags or markup changed over time? This is a concern that Twinke/AWB/etc. may have started/stopped including flags or changed what they included over time. 3. Are there other API queries or data elsewhere I could use to identify (semi-)automated revisions? On Mon, May 19, 2014 at 10:35 AM, Federico Leva (Nemo) <nemowiki(a)gmail.com -- Brian C. Keegan, Ph.D. Post-Doctoral Research Fellow, Lazer Lab College of Social Sciences and Humanities, Northeastern University Fellow, Institute for Quantitative Social Sciences, Harvard University Affiliate, Berkman Center for Internet & Society, Harvard Law School b.keegan(a)neu.edu www.brianckeegan.com M: 617.803.6971 O: 617.373.7200 Skype: bckeegan _______________________________________________ Wiki-research-l mailing list Wiki-research-l(a)lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/wiki-research-l

Another thought I had was that because many semi-automated tools such as Twinkle and AWB leave parenthetical annotations in their revision comments, would this be a relatively inexpensive way to filter out revisions rather than users? Some caveats, I'd like to get domain experts' feedback on. I'm not expecting settled research, just input from others' experiences munging the data. 1. Is the inclusion of this markup in revision comments optional? This is a concern that some users may enable or disable it, so I may end up biasing inclusion based on users' preferences.

2. How have these flags or markup changed over time? This is a concern that Twinke/AWB/etc. may have started/stopped including flags or changed what they included over time.

3. Are there other API queries or data elsewhere I could use to identify (semi-)automated revisions?

3. Are there other API queries or data elsewhere I could use to identify (semi-)automated revisions?

Thanks all for the comments on my paper, and even more thanks to everyone sharing these super helpful ideas on filtering bots: this is why I love the Wikipedia research committee. I think Oliver is definitely right that "Citation goldmine" as one friend called it, I think. This won't address edit logs to date, but do we know if most bots and automated tools use the API to make edits? If so, would it be feasibility to add a flag to each edit as to whether it came through the API or not. This won't stop determined users, but might be a nice way to identify cyborg edits from those made manually by the same user for many of the standard tools going forward. The closest thing I found in the bug tracker is [1], but it doesn't address the issue of 'what is a bot' which this thread has clearly shown is quite complex. An API-edit vs. non-API edit might be a way forward unless there are automated tools/bots that don't use the API. 1. https://bugzilla.wikimedia.org/show_bug.cgi?id=11181 Cheers, Scott _______________________________________________ Wiki-research-l mailing list Wiki-research-l(a)lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/wiki-research-l

Actually, belay that, I have a pretty good idea. I'll fire the log parser up now. On 20 May 2014 01:21, Oliver Keyes <okeyes(a)wikimedia.org> wrote: -- Oliver Keyes Research Analyst Wikimedia Foundation

Okay. Methodology: *take the last 5 days of requestlogs; *Filter them down to text/html requests as a heuristic for non-API requests; *Run them through the UA parser we use; *Exclude spiders and things which reported valid browsers; *Aggregate the user agents left; *??? *Profit It looks like there are a relatively small number of bots that browse/interact via the web - ones I can identify include WPCleaner[0], which is semi-automated, something I can't find through WP or google called "DigitalsmithsBot" (could be internal, could be external), and Hoo Bot (run by User:Hoo man). My biggest concern is DotNetWikiBot, which is a general framework that could be masking multiple underlying bots and has ~ 7.4m requests through the web interface in that time period. Obvious caveat is obvious; the edits from these tools may actually come through the API, but they're choosing to request content through the web interface for some weird reason. I don't know enough about the software behind each bot to comment on that. I can try explicitly looking for web-based edit attempts, but there would be far fewer observations that the bots might appear in, because the underlying dataset is sampled at a 1:1000 rate. [0] https://en.wikipedia.org/wiki/User:NicoV/Wikipedia_Cleaner/Documentation On 20 May 2014 07:50, Oliver Keyes <okeyes(a)wikimedia.org> wrote: -- Oliver Keyes Research Analyst Wikimedia Foundation _______________________________________________ Wiki-research-l mailing list Wiki-research-l(a)lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/wiki-research-l