Hey Trey
Thanks for the in depth discussion. So if the terms people are using that result in "zero search results" are typically gibberish why do we care if 30% of our searches result in "zero search results"? A big deal was made about this a while ago.
If one was just to look at those search terms that more than 100 IPs searched for would that not remove the concerns about anonymity? One could also limit the length of the searches displaced to 50 characters. And just provide the first 100 with an initial human review to make sure we are not miss anything.
James
On Fri, Jul 15, 2016 at 9:31 AM, Trey Jones tjones@wikimedia.org wrote:
Pine, thanks for the forward. Regulars on the Discovery list may know me, but James probably does not. I've manually reviewed tens of thousands of generally poorly performing queries (fewer than 3 results) and skimmed hundreds of thousands more from many of the top 20 Wikipedias—and to a lesser extent other projects—over the year I've been at the WMF and in Discovery. You can see my list of write ups here https://www.mediawiki.org/wiki/User:TJones_(WMF)/Notes.
So I want to say that this is an awesome idea—which is why many people have thought of it. It was apparently one of the first ideas the Discovery department had when they formed (see Dan's notes linked below). It was also one of the first ideas I had when I joined Discovery a few months later.
Dan Garry's notes on T8373 https://phabricator.wikimedia.org/T8373#1856036 and the following discussion pretty much quash the idea of automated extraction and publication from a privacy perspective. People not only divulge their own personal information, they also divulge other people's personal information. One example: some guy outside the U.S. was methodically searching long lists of real addresses in Las Vegas. I will second Dan's comments in the T8373 discussion; all kinds of personal data end up in search queries. A dump of search queries *was* provided in September 2012 https://blog.wikimedia.org/2012/09/19/what-are-readers-looking-for-wikipedia-search-data-now-available/, but had to be withdrawn over privacy concerns.
Another concern for auto-published data: never underestimate the power of random groups of bored people on the internet. 4chan decided to arrange Time Magazine poll results https://techcrunch.com/2009/04/27/time-magazine-throws-up-its-hands-as-it-gets-pwned-by-4chan/ so the first letter spelled out a weird message. It would be easy for 4chan, Reddit, and other communities to get any message they want on that list if they happened to notice that it existed. See also Boaty McBoatface https://en.wikipedia.org/wiki/RRS_Sir_David_Attenborough#Name and Mountain Dew "Diabeetus" https://storify.com/cbccommunity/hitler-did-nothing-wrong-wins-crowdsourced-mounta (which is not at all the worst thing on *that* list). We don't want to have to try to defend against that.
In my experience, the quality of what's actually there isn't that great. One of my first tasks when I joined Discovery was to look at daily lists of top 100 zero-results queries that had been gathered automatically. I was excited by this same idea. The top 100 zero-results query list was a wasteland. (Minimal notes on some of what I found are here https://www.mediawiki.org/wiki/User:TJones_(WMF)/Notes/Survey_of_Zero-Results_Queries#Highly_repeated_searches.) We could make it better by focusing on human-ish searchers, using basic bot-exclusion techniques https://www.mediawiki.org/wiki/User:TJones_(WMF)/Notes/TextCat_Optimization_for_frwiki_eswiki_itwiki_and_dewiki#Random_sampling, ignoring duplicates from the same IP, and such, but I don't think it would help. And while Wikipedia is not for children, there could be an annoying amount of explicit adult material on the list, too. We would probably find some interesting spellings of Facebook and WhatsApp, though.
If we're really excited about this, I could imagine using better techniques to pull zero-results queries and see if anything good is in there, but we'd have to commit to some sort of review before we publish it. For example, Discernatron https://discernatron.wmflabs.org/ data, after consulting with legal, is reviewed independently by two people, who then have to reconcile any discrepancies, before being made public. So I think we'd need an ongoing commitment to have at least two people under NDA who would review any list before publication. 500-600 queries takes a couple hours per person (we’ve done that for the Discernatron), so the top 100 would probably be less than an hour. I'd even be willing to help with the review (as I am for Discernatron) if we found there was something useful in there—but I'm not terribly hopeful. We'd also need more people to efficiently and effectively review queries for other languages if we wanted to extend this beyond English Wikipedia.
Finally, if this is important enough and the task gets prioritized, I'd be willing to dive back in and go through the process once and pull out the top zero-results queries, this time with basic bot exclusion and IP deduplication—which we didn't do early on because we didn't realize what a mess the data was. We could process a week or a month of data and categorize the top 100 to 500 results in terms of personal info, junk, porn, and whatever other categories we want or that bubble up from the data, and perhaps publish the non-personal-info part of the list as an example, either to persuade ourselves that this is worth pursuing, or as a clearer counter to future calls to do so. —Trey
Trey Jones Software Engineer, Discovery Wikimedia Foundation
On Fri, Jul 15, 2016 at 10:09 AM, Pine W wiki.pine@gmail.com wrote:
Forwarding
Pine ---------- Forwarded message ---------- From: "James Heilman" jmh649@gmail.com Date: Jul 15, 2016 06:33 Subject: [Wikimedia-l] Improving search (sort of) To: "Wikimedia Mailing List" wikimedia-l@lists.wikimedia.org Cc:
A while ago I requested a list of the "most frequently searched for terms for which no Wikipedia articles are returned". This would allow the community to than create redirect or new pages as appropriate and help address the "zero results rate" of about 30%.
While we are still waiting for this data I have recently come across a list of the most frequently clicked on redlinks on En WP produced by Andrew West https://en.wikipedia.org/wiki/User:West.andrew.g/Popular_redlinks Many of these can be reasonably addressed with a redirect as the issue is often capitals.
Do anyone know where things are at with respect to producing the list of most search for terms that return nothing?
-- James Heilman MD, CCFP-EM, Wikipedian
The Wikipedia Open Textbook of Medicine www.opentextbookofmedicine.com _______________________________________________ Wikimedia-l mailing list, guidelines at: https://meta.wikimedia.org/wiki/Mailing_lists/Guidelines New messages to: Wikimedia-l@lists.wikimedia.org Unsubscribe: https://lists.wikimedia.org/mailman/listinfo/wikimedia-l, mailto:wikimedia-l-request@lists.wikimedia.org?subject=unsubscribe
discovery mailing list discovery@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/discovery
On 15 July 2016 at 08:44, James Heilman jmh649@gmail.com wrote:
Thanks for the in depth discussion. So if the terms people are using that result in "zero search results" are typically gibberish why do we care if 30% of our searches result in "zero search results"? A big deal was made about this a while ago.
Good question! I originally used to say that it was my aspiration that users should never get zero results when searching Wikipedia. As a result of Trey's analysis, I don't say that any more. ;-) There are many legitimate cases where users should get zero results. However, there are still tons of examples of where giving users zero results is incorrect; "jurrasic world" was a prominent example of that.
It's still not quite right to say that *all* the terms that people use to get zero results are gibberish. There is an extremely long tail https://en.wikipedia.org/wiki/Long_tail of zero results queries that aren't gibberish, it's just that the top 100 are dominated by gibberish. This would mean we'd have to release many, many more than the top 100, which significantly increases the risk of releasing personal information.
If one was just to look at those search terms that more than 100 IPs searched for would that not remove the concerns about anonymity? One could also limit the length of the searches displaced to 50 characters. And just provide the first 100 with an initial human review to make sure we are not miss anything.
The problem with this is that there are still no guarantees. What if you saw the search query "DF198671E"? You might not think anything of it, but I would recognise it as an example of a national insurance number https://en.wikipedia.org/wiki/National_Insurance_number, the British equivalent of a social security number [1]. There's always going to be the potential that we accidentally release something sensitive when we release arbitrary user input, even if it's manually examined by humans.
So, in summary:
- The top 100 zero results queries are dominated by gibberish. - There's a long tail of zero results queries, meaning we'd have to reduce many more than the top 100. - Manually examining the top zero results queries is not a foolproof way of eliminating personal data since it's arbitrary user input.
I'm happy to answer any questions. :-)
Thanks, Dan
[1]: Don't panic, this example national insurance number is actually invalid. ;-)
The "jurrasic world" example is a good one as it was "fixed" by User:Foxj adding a redirect https://en.wikipedia.org/w/index.php?title=Jurrasic_world&action=history
Agree we would need to be careful. The chance of many different IPs all searching for "DF198671E" is low but I agree not zero and we would need to have people run the results before they are displayed.
I guess the question is how much work would it take to look at this sort of data for more examples like "jurrasic world"?
James
On Fri, Jul 15, 2016 at 10:05 AM, Dan Garry dgarry@wikimedia.org wrote:
On 15 July 2016 at 08:44, James Heilman jmh649@gmail.com wrote:
Thanks for the in depth discussion. So if the terms people are using that result in "zero search results" are typically gibberish why do we care if 30% of our searches result in "zero search results"? A big deal was made about this a while ago.
Good question! I originally used to say that it was my aspiration that users should never get zero results when searching Wikipedia. As a result of Trey's analysis, I don't say that any more. ;-) There are many legitimate cases where users should get zero results. However, there are still tons of examples of where giving users zero results is incorrect; "jurrasic world" was a prominent example of that.
It's still not quite right to say that *all* the terms that people use to get zero results are gibberish. There is an extremely long tail https://en.wikipedia.org/wiki/Long_tail of zero results queries that aren't gibberish, it's just that the top 100 are dominated by gibberish. This would mean we'd have to release many, many more than the top 100, which significantly increases the risk of releasing personal information.
If one was just to look at those search terms that more than 100 IPs searched for would that not remove the concerns about anonymity? One
could
also limit the length of the searches displaced to 50 characters. And
just
provide the first 100 with an initial human review to make sure we are
not
miss anything.
The problem with this is that there are still no guarantees. What if you saw the search query "DF198671E"? You might not think anything of it, but I would recognise it as an example of a national insurance number https://en.wikipedia.org/wiki/National_Insurance_number, the British equivalent of a social security number [1]. There's always going to be the potential that we accidentally release something sensitive when we release arbitrary user input, even if it's manually examined by humans.
So, in summary:
- The top 100 zero results queries are dominated by gibberish.
- There's a long tail of zero results queries, meaning we'd have to
reduce many more than the top 100.
- Manually examining the top zero results queries is not a foolproof way
of eliminating personal data since it's arbitrary user input.
I'm happy to answer any questions. :-)
Thanks, Dan
[1]: Don't panic, this example national insurance number is actually invalid. ;-)
-- Dan Garry Lead Product Manager, Discovery Wikimedia Foundation _______________________________________________ Wikimedia-l mailing list, guidelines at: https://meta.wikimedia.org/wiki/Mailing_lists/Guidelines New messages to: Wikimedia-l@lists.wikimedia.org Unsubscribe: https://lists.wikimedia.org/mailman/listinfo/wikimedia-l, mailto:wikimedia-l-request@lists.wikimedia.org?subject=unsubscribe
Forwarded at the request of Trey Jones
Hey James,
When we first started looking at zero results rate (ZRR), it was an easy metric to calculate, and it was surprisingly high. We still look at ZRR https://searchdata.wmflabs.org/metrics/#failure_rate because it is so easy to measure, and anything that improves it is probably a net positive (note the big dip when the new completion suggester was deployed!!), but we have more complex metrics that we prefer. There's user engagement https://searchdata.wmflabs.org/metrics/#kpi_augmented_clickthroughs/augmented clickthroughs, which combines clicks and dwell time and other user activity. We also use historical click data in a metric that improves when we move clicked-on results higher in the results list, which we use with the Relevance Forge https://gerrit.wikimedia.org/r/#/admin/projects/wikimedia/discovery/relevanceForge .
And I didn't mean to give the impression that *most* zero-results queries are gibberish, though many, many are. And that was something we didn't really know a year ago. There are also non-gibberish results that correctly get zero results, like most DOI https://www.mediawiki.org/wiki/User:TJones_(WMF)/Notes/Survey_of_Zero-Results_Queries#DOI and many media player https://www.mediawiki.org/wiki/User:TJones_(WMF)/Notes/Survey_of_Zero-Results_Queries#TV_Episodes_.2F_Movies.E2.80.94.22....22_film queries. We also see a lot of non-notable (not-yet-notable?) public figures (local bands, online artists, youtube musicians), and sometimes just random names.
The discussion in response to Dan's original comment in Phab mentions some approaches to reduce the risk of automatically releasing private info, but I still take an absolute stand against unreviewed release. If I can get a few hundred people to click on a link like this https://en.wikipedia.org/w/index.php?title=Special:Search&profile=default&fulltext=Search&search=%22James+is+a+nice+guy%22, I can get any message I want on that list. (Curious? Did you click?) The message could be less anonymous and much more obnoxious, obviously.
50 character limits won't stop emails and phone numbers from making the list (which invites spam and cranks). Those can be filtered, but not perfectly.
I've only looked at these top lists by day in the past, but on that time scale the top results are usually under 1000 count (and that includes IP duplicates), so the list of queries with 100 IPs might also be very small.
As I said, I'm happy to do the data slogging to try this in a better fashion if this task is prioritized, and I'd be happy to be wrong about the quality of the results, but I'm still not hopeful.
—Trey
Trey Jones Software Engineer, Discovery Wikimedia Foundation
On Fri, Jul 15, 2016 at 10:19 AM, James Heilman jmh649@gmail.com wrote:
The "jurrasic world" example is a good one as it was "fixed" by User:Foxj adding a redirect https://en.wikipedia.org/w/index.php?title=Jurrasic_world&action=history
Agree we would need to be careful. The chance of many different IPs all searching for "DF198671E" is low but I agree not zero and we would need to have people run the results before they are displayed.
I guess the question is how much work would it take to look at this sort of data for more examples like "jurrasic world"?
James
On Fri, Jul 15, 2016 at 10:05 AM, Dan Garry dgarry@wikimedia.org wrote:
On 15 July 2016 at 08:44, James Heilman jmh649@gmail.com wrote:
Thanks for the in depth discussion. So if the terms people are using
that
result in "zero search results" are typically gibberish why do we care
if
30% of our searches result in "zero search results"? A big deal was made about this a while ago.
Good question! I originally used to say that it was my aspiration that users should never get zero results when searching Wikipedia. As a result of Trey's analysis, I don't say that any more. ;-) There are many legitimate cases where users should get zero results. However, there are still tons of examples of where giving users zero results is incorrect; "jurrasic world" was a prominent example of that.
It's still not quite right to say that *all* the terms that people use to get zero results are gibberish. There is an extremely long tail https://en.wikipedia.org/wiki/Long_tail of zero results queries that aren't gibberish, it's just that the top 100 are dominated by gibberish. This would mean we'd have to release many, many more than the top 100, which significantly increases the risk of releasing personal information.
If one was just to look at those search terms that more than 100 IPs searched for would that not remove the concerns about anonymity? One
could
also limit the length of the searches displaced to 50 characters. And
just
provide the first 100 with an initial human review to make sure we are
not
miss anything.
The problem with this is that there are still no guarantees. What if you saw the search query "DF198671E"? You might not think anything of it, but I would recognise it as an example of a national insurance number https://en.wikipedia.org/wiki/National_Insurance_number, the British equivalent of a social security number [1]. There's always going to be the potential that we accidentally release something sensitive when we release arbitrary user input, even if it's manually examined by humans.
So, in summary:
- The top 100 zero results queries are dominated by gibberish.
- There's a long tail of zero results queries, meaning we'd have to
reduce many more than the top 100.
- Manually examining the top zero results queries is not a foolproof
way of eliminating personal data since it's arbitrary user input.
I'm happy to answer any questions. :-)
Thanks, Dan
[1]: Don't panic, this example national insurance number is actually invalid. ;-)
-- Dan Garry Lead Product Manager, Discovery Wikimedia Foundation _______________________________________________ Wikimedia-l mailing list, guidelines at: https://meta.wikimedia.org/wiki/Mailing_lists/Guidelines New messages to: Wikimedia-l@lists.wikimedia.org Unsubscribe: https://lists.wikimedia.org/mailman/listinfo/wikimedia-l, mailto:wikimedia-l-request@lists.wikimedia.org?subject=unsubscribe
-- James Heilman MD, CCFP-EM, Wikipedian
The Wikipedia Open Textbook of Medicine www.opentextbookofmedicine.com
If I can get a
few hundred people to click on a link like this https://en.wikipedia.org/w/index.php?title=Special:Search&profile=default&fulltext=Search&search=%22James+is+a+nice+guy%22, I can get any message I want on that list. (Curious? Did you click?) The message could be less anonymous and much more obnoxious, obviously
They could vandalize any one of over ten million pages on the English Wikipedia and get the same result. We should be conscious of the dangers but we can easily route around them like we do with other kinds of vandalism.
How hard would it be to ask for search feedback on search results, perhaps piloting with some small subset of zero-result searches? For 1/1000 ZRRs, prompt the user to provide some type of useful information about why there should be results, or if there ought to be, or what category of information the searcher was looking for, etc. You'd get junk and noise, but it might be one way to filter out a lot of the gibberish. You could also ask people to agree to make their failed search part of a publicly visible list, although this could of course be gamed.
wikimedia-l@lists.wikimedia.org