Skip to content

Mix’n’match post-mortem

So this, as they say, happened.

On 2016-12-27, I received an update on a Mix’n’match catalog that someone had uploaded. That update had improved names and descriptions for the catalog. I try to avoid such updates, because I made the import function so I do not have to deal with every catalog myself, and also because the update process is entirely manual, therefore somewhat painful and error-prone, as we will see. Now, as I was on vacation, I was naturally in a hurry, and (as it turned out later) there were too many tabs in the tab-delimited update file.

Long story short, something went wrong with the update. For some reason, some of the SQL commands I generated from the update file did not specify some details about which entry to update. Like, its ID, or the catalog. So when I checked what was taking so long, just short of 100% of Mix’n’match entries had the label “Kelvinator stove fault codes”, and the description “0”.

Backups, you say? Well, of course, but, look over there! /me runs for the hills

Well, not all was lost. Some of the large catalogs were still around from my original import. Also, my scraping scripts for specific catalogs generate JSON files with the data to import, and those are still around as well. There was also a SQL dump from 2015. That was a start.

Of course, I did not keep the catalogs imported through my web tool. Because they were safely stored in the database, you know? What could possibly go wrong? Thankfully, some people still had their original files around and gave them to me for updating the labels.

I also wrote a “re-scraping” script, which uses the external URLs I store for each entry in Mix’n’match, together with the external ID. Essentially, I get the respective web page, and write a few lines of code to parse the <title> tag, which often includes the label. This works for most catalogs.

So, at the time of writing, over 82% of labels in Mix’n’match have been successfully restored. That’s the good news.

The bad news is that the remaining ~17% are distributed across 133 catalogs. Some of these do not have URLs to scrape, some URLs don’t play nicely (session-based Java horrors, JS-only pages etc.), and the rest need site-specific <title> scraping code. Fixing those will take some time.

Apart from that, I fixed up a few things:

  • Database snapshots (SQL dump) will now be taken once a week
  • The snapshot from the previous week is preserved as well, in case damage went unnoticed
  • Catalogs that are uploaded through the import tool will be preserved as individual files

Other than the remaining entries that require fixing, Mix’n’match is open for business, and while my one-man-show is spread thin as usual, subsequent blunders should be easier to mitigate. Apologies for the inconvenience, and all that.

All your locations are belong to us

A recent push for a UK photography contest reminded me of an issue I have begrudged for a quite a while. On the talk page for that contest, I pointed to several tools of mine, dealing with images and locations. But they only show aspects of those, like “Wikidata items without images”. What about the others? WDQS can show maps of all Wikidata items in a region, but what about Wikipedia? The mobile app can show you things with Wikipedia articles nearby, but what about Commons? I don’t recall a way to see Commons images taken near a location (WD-FIST can find them, but without a map). The data exists, but is either hard to get to, or “siloed” in some tool/app.

Screen Shot 2016-08-20 at 18.06.26Wouldn’t it be great to get a map with all this information on it? All of Wikidata? All of Wikipedia? All of Commons? At once?

What should that look like? The photography contest scenario, and change in general web usage patterns, suggest a strong emphasis on mobile. Which in turn tends to be “no frills”, as in, a focus on what is important: The map, and the objects on it.

So I decided (for the time being) to get rid of the query functions in WD-FIST, and the clutter in WikiShootMe, and start from scratch, with (essentially) just a big map, using the bleeding-edge versions of JS libraries like bootstrap, jQuery, and leaflet. So without further ado, I present WikiShootMe, version 3 (pre-alpha). As it is, the tool defaults to your coordinates, which may be your local hub (as in my case, in the screenshot). There are four layers, which can be individually toggled:

  1. Wikidata items with images (in green)
  2. Wikidata items without image (in red, the Wikipedia will change with your language selection)
  3. Commons images (in blue)
  4. Wikipedia articles (smaller, in yellow, mostly overlapping Wikidata items)

There is also a grey circle in the center, which is your (or your local hub’s) position. On mobile, this should move with you (but I haven’t tested that, as it would require leaving the house). All of these have a pop-up, when you click or touch the circle. It shows the linked title of the object, and, for Wikidata items with images and Commons images, it shows the respective image.

All these data sources will update when you move the map, as well as zoom, up to a certain zoom factor. Below that, an “Update” button will appear to update manually, but it can take a long time, even with the number of objects limited.

Screen Shot 2016-08-20 at 18.46.05I find it amazing how many geo-coded images there are already on Commons (even though the API will only give me 500 at a time). Maybe that is the geograph effect here in the UK, which let to the import of hundreds of thousands of free images to Commons.But I also found a funny pattern in Cologne, Germany, which turned out to be a series of images taken by Wikimedia volunteers from a balloon!

Now, to be extra clever, I tried to add an upload function to the pop-up of Wikidata items without an image. You can select a file from disk, or use the camera as a source on mobile. It will pre-fill the title and the {{Information}} template with a link to the respective Wikidata object. However, several problems occur with that:

  • I only could get the “old” Commons upload page to work with the pre-filled data
  • I could find no documentation on <form> parameters for the Upload wizard
  • I haven’t actually tested if the upload works
  • There seems to be no way to automatically add the uploaded image to the Wikidata item

A way around all that would be to upload the image to the tool itself, then transfer it to Commons via OAuth. This would also allow me to add the new image as the P18 on the Wikidata item. This is an option to be explored, especially if the Upload Wizard remains opaque to me.

Update: I have added OAuth to the tool. Once authorised, you can upload a new image for a Wikidata item from both desktop and mobile (gallery or camera directly) with one click. It fills in file name, coordinates, default license etc. It even adds the image to the item after upload automatically. All this opens in a new tab, on the page for the uploaded image, to give you a chance to add more information.

As usual, I am quite open for bug reports, feature requests (yes, it’s bare-bones at the moment), and technical support by volunteers/WMF.

Livin’ on the edge

A few days ago, Lydia posted about the first prototype of the new structured data system for Commons, based on Wikidata technology. While this is just a first step, structured data for Commons seems finally within reach.

And that brings home the reality of over 32 million files on Commons, all having unstructured data about them, in the shape of the file description pages. It would be an enormous task do manually transcribe all these descriptions, licenses, etc. to the appropriate data structures. And while we will have to do just that for many of the files, the ones that can be transcribed by a machine, should be.

So I went ahead and re-wrote a prototype tool I had build for just this occasion a while ago. I call it CommonsEdge (a play on Common sedge). It is both an API, and an interface to that API. It will parse a file description page on Commons, and return a JSON object with the data elements corresponding to the description page. An important detail is that this parser does not just pick some elements it understands, and ignore the rest; internally, it tries to “explain” all elements of the description (templates, links, categories, etc.) as data, and fails if it can not explain one. That’s right, the API call will fail with an error, unless 100% of the page would be represented in the JSON object returned. This prevents “half-parsed” pages; a file description page that is successfully pared by the API can safely be replaced in its entirety by the resulting structured data. In case of failure, the error message is usually quite specific and detailed about the cause; this allows for incremental improvements of the parser.

Screen Shot 2016-08-03 at 21.35.19At the moment of writing, I find that ~50-60% of file descriptions (based on sets of 1000 random files) produce a JSON object, that is, can be completely understood by the parser, and completely represented in the result. That’s 16-19 million files descriptions that can be converted to structured data automatically, today. Most of the failures appear to be due to bespoke templates; the more common ones can be added over time.

A word about the output: Since the structured data setup, including properties and foreign keys, is still in flux, I opted for a simple output format. It is not Wikibase format, but similar; most elements (except categories and coordinates, I think) are just lists of type-and-value tuples (example). I try to use URLs as much as possible, for example, when referencing users on Commons (or other Wikimedia projects) or flickr. Licenses are currently links to the Wikidata element corresponding to the used template (ideally, I would like to resolve that through Wikidata properties pointing to the appropriate license).

Source code is available. Pull requests are welcome.

WDQ, obsolete?

Since a few years, I run the WikiData Query tool (WDQ) to provide a query functionality to Wikidata. Nowadays, the (confusingly similarly named) SPARQL-based WDQS is the “official” way to query Wikidata. WDQS has been improving a lot, and while some of my tools still support WDQ, I deliberately left that option out of new tools like PetScan. But before I shut down WDQ, and the tools that use it, for good, I wanted to know if it is still used, and if SPARQL could take over.

I therefore added a query logger to Autolist1 and Autolist2. The logs contain all WDQ queries run through those tools. I will monitor the results for a while, but here is what I saw so far. I will comment on translating the query to SPARQL using WDQ2SPARQL, the general ability for such queries, and the performance of WDQS. “OK” means the query could be converted automatically to SPARQL, runs, and produces a similar (as in, equal or more up-to-date) result.

WDQ Comment
CLAIM[279:13219666]  OK
BETWEEN[569,1016-1,1016-12] BETWEEN not implemented in WDQS, but manual translation feasible

Update: This has been implemented by smalyshev no, runs OK!

(CLAIM[1435:10387684] OR CLAIM[1435:10387575]) AND NOCLAIM[380] AND NOCLAIM[481]  OK
BETWEEN[569,1359-1,1359-12] BETWEEN not implemented in WDQS, but manual translation feasible

Update: This has been implemented by smalyshev no, runs OK!

CLAIM[31:5]  All humans ~3.2M humans on Wikidata. Not really a useful query in these tools.
Q22686  Single item. Doesn’t really need a query?
Q22686  Single item. Doesn’t really need a query?
CLAIM[106:170790] AND CLAIM[27:35]  OK
CLAIM[195:842858]  OK
Gustav III  What the hell?
claim[17]  All items with “country”. Not really a useful query in these tools.
claim[31]  All items with “instance of”. Not really a useful query in these tools.
claim[106:82955] and claim[509:(tree[12078][][279])]  OK
claim[31:5]   All humans ~3.2M humans on Wikidata. Not really a useful query in these tools.
claim[31:5]   All humans ~3.2M humans on Wikidata. Not really a useful query in these tools.
claim[21]  All items with gender. Not really a useful query in these tools.
LINK[lvwiki] AND CLAIM[31:5]  OK
LINK[lvwiki] AND CLAIM[31:5]  OK
claim[27] and noclaim[21]  OK
LINK[lvwiki] AND CLAIM[31:56061]  OK
LINK[lvwiki] AND tree[56061][150][17,279]  OK
claim[31:(tree[16521][][279])]  OK

As far as I can tell, SPARQL could take over for WDQ immediately.


Wikipedia has language editions, Wikidata has labels, aliases, descriptions, and some properties in multiple languages. This a great resource, to get the world’s knowledge in your language! But looking at the technical site, things become a little dim. Wikimedia sites have their interface translated in many languages, but beyond that, English rules supreme. Despite many requests, only few tools on Labs have a translatable (and translated) interface.

One exception is PetScan, which uses the i18n mechanism from its predecessor CatScan, namely a single wiki page on meta, which contains all interface translations. This works in principle, as the many translations there show, but it has several disadvantages, ranging from bespoke wikitext parsing, over load/rendering times on meta, to the fact that there is no easy way to answer the question “which of these keys have not been translated into Italian”? New software features require new interface strings, so the situation gets worse over time.

The answer I got when asking about good ways to translate interfaces is usually “just use TranslateWiki“, which IIRC is used for the official Wikimedia sites. This is a great project, with powerful applications, but I was looking for something more light-weight, both on the “add a translation” side, and the “how to use this in my tool” side.

ToolTranslateIf you know me or my blog, then by this point, you will already have guessed what happened next: I rolled my own (for more detailed information, see the manual page).

ToolTranslate is a tool that allows everyone (after the usual OAuth ceremony) to provide translations for interface texts, in almost 300 languages. I even made a video demonstrating how easy it is to add translations (ToolTranslate uses its own mechanism, so the demo edit shows up live in the interface). You can even also your own tool, without having to jump through bureaucratic hurdles, just with the press of a button!

On the tool-author side, you will have to change your HTML, from <div>My text</div> to <div tt=”mytext”></div>, and then add “My text” as a translation for the “mytext” key. Just use the language(s) you know, anyone can add translations in other languages later.I experienced this myself; after I uploaded the demo video, User:Geraki added Greek translations to the interface, before this blog post, or any other instructions, were available. Just, suddenly, as if by magic, Greek appeared as an interface option… You will also need to include a JavaScript file I provide, and add a single line of code (two, if you want to have a drop-down to switch languages live).

There is a simplistic demo page, mainly intended for tool authors, to see how it works in practice. Besides ToolTranslate itself, I also used it on my WikiLovesMonuments tool, to show that it is feasible to retrofit an existing tool. This took less than 10 minutes.

I do provide the necessary JavaScript code to convert HTML/JS-based tools. I will be working on a PHP class next, if there is demand. All translations are also provided as JSON files online, so you can, in turn, “roll your own” code if you want. And if you have existing translations for your tool and want to switch to ToolTranslate, let me know, and I can import your existing translations.

First image, good image?

For a while now, Wikimedia pages (usually, Wikipedia articles) have a “page image”, an image from that page used as a thumbnail in article previews, e.g. in the mobile app. While it is not entirely clear to me how this is image is chosen, it appears to be the first image of the article in most cases, probably excluding some icons.

Wikidata is doing something similar with the “image” property (P18), however, this needs to be an image of the item’s subject, not “something related to the item”. Wikipedia’s “page image” often turns out to be a painting made by the article’s subject, or a map, or something related to an event. This discrepancy prevent an automated import of the “page image” into Wikidata. However, exceptions aside, the “page item” presents a highly specific resource for P18-suitable images.

Screen Shot 2016-07-18 at 10.38.02So I added a new function to my WD_FIST tool, to help facilitate the import of suitable images from that rich source into Wikidata. As a first step, a bot checks several large Wikipedias on a daily basis, and retrieves “page images” where the associated Wikidata item has none, and the “page image” is stored on Commons. It also skips “non-subject” pages like list articles. In a second stage, images (excluding PNG, GIF, and SVG) that are used as a “page image” on at least three Wikipedias for the same subject are put into a main candidate list. The image must also not be on the tool-internal “ignore” list. Even after all this filtering, >32K candidates remain in the current list.

dewiki 346,204
enwiki 700,832
frwiki 255,527
itwiki 148,041
nowiki 73,508
plwiki 181,323
svwiki 109,349
Combined 32,137

I will likely add more Wikipedias to this list (es and pt will show up tomorrow), and eventually lower the inclusion threshold, as candidates are added to Wikidata, or to the “ignore” list.

As the candidate list is already heavily filtered, I am not applying some of the usual WD-FIST filters. This also helps with retrieving a candidate set of 50 very quickly. In this mode, the tool also lends itself well to mobile usage.

A week of looking at women

Images and their use in the WikiVerse have always been a particular interest of mine, on Wikipedia, Commons, and of course, Wikidata. Commons holds the files and groups them by subject, author, or theme; Wikidata references images and files for key aspects of a subject; and Wikipedia uses them to enrich texts, and puts files into context.

Wikidata uses images for more subjects than any Wikipedia, save English, and it is slowly encroaching on the latter; the “break-even” should happen later this year. This is not just a purpose in itself, but will also massively benefit the many smaller Wikipedias, by holding such material in easily usable form at the ready.

Screen Shot 2016-04-07 at 00.10.26

Image candidates, ready to be added with a single click

So I did a small experiment, as to how much one person can do “on the side” (besides work, other interests, and such luxuries as sleeping or eating), to improve the Wikidata image fundus. I thus picked the German category for women, which currently holds >92K articles. I used my WD-FIST tool to find all potential images on all Wikipedias, for the Wikidata items corresponding to the German articles. This does not show items that already have an image, or items that have no possible candidate image anywhere; just the ones where a Wikipedia does have an image, and Wikidata does not.

A week ago, I started with 3,060 items of women that potentially had an image on Wikipedia, somewhere. A week later, I am down to ~290. Now, that does not mean I added ~2,700 images to Wikidata; a database query comes to about ~1,100 added images, and ~200 other file properties (spoken text, commemorative plaque image, etc.). Some items just had no suitable image on Wikipedia; others had group photos, which I tagged to be cropped on Commons (those tagged images will not show in the tool, while the crop template remains).

The image candidates for the remaining 290 or so items need to be investigated in more detail; some of them might be not actually images of the subject (hard to tell if the file name and description are in e.g. Russian), or they are low-resolution group pictures, which do not warrant cropping, as the resulting, individual image would be too grainy.

Adding the ~1,100 images is good, but only part of the point I am trying to make here. The other part is, no one will have to wade again through the ~90% of item/image suggestions I have resolved, one way or another. Ideally, the remaining 290 items should be resolved to, so if an image is added on any Wikipedia, for any of the >92K women in the category, just that new image would show in the tool, which would make updating Wikidata so much easier. Even just one volunteer could drop by every few weeks and keep Wikidata up-to-date with images, for that group of items, with a few clicks’ worth.

The next step is, of course, all women on Wikidata (caution: that one will load a few minutes). The count of items with potential images is at 15,986 at the time of writing. At my speed, it would take one person about a month of late evening clicking to reduce that by 90%, though I do hope some of you have been inspired to help me out a bit.

Of cats and pets

CatScan is one of these workhorse tools that are familiar to many Wikimedia users, all the way back to the toolserver. Its popularity, however, has also caused problems with reliability time and again. As Labs became usable, I added QuickIntersection to the mix, allowing for a quicker and more reliable service at the expense of some complex functionality. Alas, despite my best efforts, CatScan reliability is fluctuating a lot. The reasons for that include the choice of PHP as a programming language, and the shared nature of Labs tools, where resources are concerned.

So I spent the last two weeks (as time allowed) with a complete rewrite of the tool, using C++ and a dedicated virtual machine on Labs. The result is one of the most complex tools I developed to date.I call it PetScan, both to indicate that it does more than just cat(egorie)s, and as a pun on the more versatile PET scan (compared to the CAT scan).

Its basic interface is based on CatScan, and it is backwards-compatible for both URL parameters (so if you have a CatScan URL, you just need to replace the server name) and output (so the JSON output will be almost identical). It can also be switched to QuickIntersection output with a parameter, so it could replace that tool as well.

But PetScan is much more encompassing. Several times before, I tried to “connect” my (and other) tools, the last time via PagePile; however, the uptake was rather low. It is clear that most users prefer a tool that slices and dices. This is why PetScan can also process other data sources, like the Wikidata SPARQL query, manual page or item lists, and yes, PagePile. Given more than one source, it builds a subset of the respective results, even if they are on other wikis (via Wikidata).

You want a list of all cats known to Wikidata that are also in the category tree of the battleship “Bismarck” on English Wikipedia? No problem. You can chose which of the input wikis should be the output wiki, so you can have the same result as Wikidata items. Now, for the latter, you might have seen an additional box at the top of the results; this is the full functionality from AutoList 2, directly available on your resulting items.

Additional goodies include:

  • Interface language can be switched “live”. The translations were copied from the CatScan translations, so that effort is re-used.
  • Namespaces are updated live when you change the wiki for the categories
  • Both templates and (incoming) links can now be a primary source, instead of just being filters for categories
  • You can filter the results by a regular expression. This works on page titles, or Wikidata labels, respectively
  • For Wikidata results, you can specify the label language used (will default to the interface language)
  • Show only Wikipedia pages without a Wikidata item
  • Only the first 10K results will be shown in HTML mode, as to not crash your browser. Others (e.g. JSON) will get you all results

I have tested PetScan on my own, but with a project of this complexity, bugs will only become apparent with many users, over time, so please help testing it. Eventually, I believe this tool can (and will) replace CatScan, QuickIntersections, Autolist, and maybe others as well.

The Reference Wars

In a recent Wikipedia Signpost Op-Ed, Andreas Kolbe wrote about Wikidata and references. He comes to the conclusion that Wikidata needs more (non-Wikipedia) references, a statement I wholeheartedly agree with. He also divines that this will never happen, that Wikidata is doomed, while at the same time somehow being controlled by Google and Microsoft; I will not comment on these “conclusions”, as others have already done so elsewhere.

Andreas also uses my own Wikidata statistics to make his point about missing references on Wikidata. The numbers I show are useful, IMHO, to show the remarkable progress of Wikidata, but they are much too crude to draw conclusions about the state of references there. Also, the impression I get from Andreas’ text is that, while Wikipedia has some issues, references are basically OK, whereas they are essentially non-existent in Wikidata.

So I thought I’d have a look at some actual numbers, especially comparing Wikipedia and Wikidata in terms of references.

One key issue is that there is no build-in way to get metrics about statements and references from Wikipedia. I therefore developed my own approach. Given a Wikipedia article, I use the REST API to get HTML for the article. I then count the number of reference uses (essentially, <ref> tags) in the article; note that this number is larger then (or at least equal to) the number of references at the bottom of the page. Then, I strip the HTML tags, and count the number of sentences (starts with an upper-case character, has at least 50 characters, ends with a “.”); the numbers were confirmed manually for a few example articles through other sentence counting tools on the web, and yielded similar results. I then assume that each sentence in the article contains one statement (or fact); in reality, there are likely many such statements (such as the first sentence of a biographical article), but I am aiming for a lower boundary here. (Any sentence not containing a statement/fact should be deleted from Wikipedia anyway.) A useful metric from both the number of reference uses, and the number of statements (=sentences), is the references-per-statement (RPS) ratio.

For Wikidata, a similar metric can be calculated. For practical purposes, I skip statements of the “string” type, as they are mostly external references in themselves (e.g. VIAF identifiers); I also skip “media”-type statements, as they should have “references” in their file description page on Commons. For references, I do not count “imported from Wikipedia”, as these are not “real” references, but rather placeholders for future improvement. Again, a RPS ratio can be computed.

I then calculated these ratios for 4,683 Featured Articles from English Wikipedia and their associated Wikidata items (data). As these articles have been significantly worked over and approved by the English Wikipedia community, they should represent the “best case scenario” for Wikipedia.

Indeed, the RPS ratio is higher for Wikipedia in 87% of cases, which would mean that Wikipedia is better referenced than Wikidata. But keep in mind that this represents the best of the best of the best of English Wikipedia articles, fifteen years in the making, compared to a three-and-a-half-year old Wikidata (and references were not supported for the first year or so). This is as good as it gets for Wikipedia, and still, Wikidata has a better RPS in about 13% of cases.

Even more interesting IMHO: Taking the mean of both number of statements and number of references for both Wikipedia and Wikidata, respectively, and calculating the RPS ratios for those means, yield 0.32 for Wikipedia and 0.15 for Wikidata. This seems counter-intuitive, given the previous 87/13 “ratio of ratios”. However, further investigation shows that only 1305 (~28%) of Wikidata items have any references at all, but where there are references, they usually outshine Wikipedia; about half of the items with at least one reference have a better RPS ratio than the respective Wikipedia article. This seems to indicate a “care factor” at work; where someone cared about adding references to the item, it was done quite well. Wikidata RPS ratios range up to 1.5, meaning two statements are, on average, supported by three references, whereas Wikipedia reaches “peak RPS ratio” at 0.93, or slightly less than one reference per statement.

I believe these numbers show that Wikidata can equal and surpass Wikipedia in terms of “referencedness”, but it is a function of attention to the items. Which in turn is a matter of man- and bot-hours spent. Indeed, for the Wikidata showcase items (the equivalent of Featured Articles on Wikipedia), the Wikidata RPS ratio is better that that of the associated English Wikipedia article in 19 out of 24 cases (~80%).

So will Wikidata ever catch up to Wikipedia in terms of RPS ratio? I think so. The ability of Wikidata to be reliably edited by a machine allows for improvement by automated and semi-automated bots, tools, games, on-wiki gadgets, etc. which allow for much steeper editing rate, as I demonstrated previously for images, where Wikidata went from nothing to second place in about two years, and is now angling for the pole position (~1.1M images at the moment). I see no reason to doubt this will happen to references as well.

The beatings will continue until morale improves

So over the weekend, Wikimedia Labs ran into a bit of trouble. Database replication broke, and was lagging about two days behind the live databases. But, thanks to tireless efforts by JCrespo, replication has now picked up again, and replication lag should be back to normal soon (even though there might be a few bits missing).

Now, this in itself is not something I would blog about; things break, things get fixed, life goes on. But then, I saw a comment by JCrespo with a preliminary analysis of what happened, and how to avoid it happening again:

“…it is due to the contraints we have for labs in terms of hardware and human resources. In order to prevent this in the future, I would like to discuss enforcing stronger constraints per user/tool.”

So, there are insufficient resources invested into (Tools) Labs. The solution, obviously, is to curtail the use of resources. This train of thought should be familiar to everyone whose country went to a phase of austerity in recent years. Even though, it now seems to be commonly agreed outside the cloudy realm of politicians, that austerity is the wrong way to go. If you have a good thing going, and you require some more resources to keep it that way, you give it more resources. You do not cut away scarce resources even more! This is how you go the way of Greece.

This is how you go the way of the toolserver.