This is an archive of past discussions. Do not edit the contents of this page. If you wish to start a new discussion or revive an old one, please do so on the current talk page.
Can we allow to provide a tag if suggested tags are too general? For example I was processing photograph of Obama and the suggestion was "male", etc. I would like to ignore suggested tags and manually add item for Barack Obama (Q76). --Jarekt (talk) 20:32, 12 December 2019 (UTC)
Yes, that would be great if the development team will change their mind and add it. Othewise I am affraid images will be flooded by less specific tags and it will demage the yield we can have from Structured data.--Juandev (talk) 11:31, 13 December 2019 (UTC)
2M16-S-0095-59-67 Två biskopar och en missionär. Helge Fosseus, Sven Danell och troligen missionär Majken Johansson
:+1 agree looks like the training data is baised on the Catolish church and all bishops are suggested to be Catolish bishops... Are we getting an AI biased Wikipedia? example picture were it was suggested we saw
Then I tried tagging lady. It turns out the tag is lady (Q1378024), which is obviously not a suitable value for P180. woman (Q467) would be a proper choice, but currently neither its label nor aliases contain lady. (I'm adding it.)--Roy17 (talk) 16:41, 14 December 2019 (UTC)
Overly broad suggestions
It feels like the suggestions are so broad that adding any of them will just create a big mess so that the depicts will be of even less use than the categories on Commons. For example, for the image File:Västra_Ingelstads_kyrka_2010_-_01.JPG it gives the suggestions chapel architecture parish parish church building landmark steeple sky spire. It already has the category Category:Västra Ingelstads kyrka which is connected to Västra Ingelstad Church (Q10718275) and would be a proper suggestion instead of half of the ones that were given. If the best depicts cannot be given as a suggestion even when they already are connected through categories I almost lean towards suggesting this tool be disabled as it will create more mess than value. Ainali (talk) 18:37, 14 December 2019 (UTC)
Actually, is the idea that the tages will be SO generic? If there is a church tag, then we probably have dozens of millions photographs of churches; should the name of the church be a better tag? Or both?--Ymblanter (talk) 09:19, 15 December 2019 (UTC)
Thats a good point. If we wont improve this toll it can mess up Commons completly. We really need to cooperate with Google to train their AI better and ad the possibility to edit our own tags. --Juandev (talk) 01:47, 17 December 2019 (UTC)
At Europeana in Lissabon Harald Sack Professor for Information Services Engineering, spoke about the importance to use metadata when doing AI picture recognition and had Wikidata as example I feel this prototype must use better categories, GPS location, Wikidata when suggesting tags....
Hi, I only enabled this feature yesterday. First problem is, I cannot add a tag myself. Well that might be down to how you want this thing to work. Second problem is, I don't have a link to the tagging screen. Today I got a tagging review request in my notifications. I have clicked that notification. Thank's to the wonders of the notification system, it is now lost forever. So I'd rather have a link of some sort to tags than a bookmark.
So I click into the tag review screen and it gives me File:Susan Anderson, Mount Pisgah Cemetery.JPG, which I edited once, to start. I have only 3 options. Publish, reset, and skip. Only I don't have those options. The only button working is skip. I'm not sure how pressing skip skip skip is going to be a useful contribution. ~ R.T.G11:49, 16 December 2019 (UTC)
You need to click on tags, and once you mark one of the tags you enable save. I can not save though, I keep getting a technical error.--Ymblanter (talk) 12:22, 16 December 2019 (UTC)
The tool does not work. After tagging and confirming the tags, I always get: Something went wrong, please try again later. --Clic (talk) 07:27, 7 February 2020 (UTC)
I tried the tool today for the first time, but it seems the edits do not happen, even though the system accepts my "save". If things went right, I would see the tags under the "structured data" tab, right? --Kritzolina (talk) 17:50, 10 February 2020 (UTC)
I had previously filled out the "Structured Data" card by hand. Saving the "tags" also worked, but they are nowhere to be found. --Georgfotoart (talk) 18:14, 10 February 2020 (UTC)
Thanks! How can I find out what an image is tagged with already then? Also - I do not see those tags I added as edits anywhere. I thought this was supposed to show up as normal edits? I am back to adding by hand, yes, I had done this before, too and thought I had found a better process, but so far it has been not really helpful. --Kritzolina (talk) 19:20, 10 February 2020 (UTC)
I can only choose the proposed tags but not add any other tag. So this tool is not working for me (I use Firefox latest version). --Hadi (talk) 17:30, 12 February 2020 (UTC)
Temporarily pausing the hold on images before tags are suggested
Currently, the tool is set to hold images for labeling for 48 hours before notifying users. This configuration is showing some problems in the system that handles it, and the team will be out of office for the most part next week for seasonal holidays going into the new year. So the temporary hold is going to be disabled until the team can fix it when they get back; users will be notified almost instantaneously of uploads for tagging. This temporary change should only last through January, then the tool will go back to holding images for processing. Keegan (WMF) (talk) 18:46, 19 December 2019 (UTC)
Firstly, these are not "tags". I might tag an picture of, say, an autumnal tree with "gold"; but that's not what is depicted.
Many of the "tags" offered (for example "stock photography", "monochrome") are concepts that are unlikely ever to be suitable for a "depicts" statement. Suitable filtering should be applied.
@Pigsonthewing: Thanks for the feedback. Some of the features you suggest have been mentioned by others on this page (adding one's own depicts) and some ideas are new (perhaps do something about showing images already repeatedly skipped), so this is much appreciated to support feature requests as well as looking at some behavior that might not be desired. I've passed it along to the development team, they'll see where this all fits in with plans for improving the tool. Keegan (WMF) (talk) 21:49, 11 February 2020 (UTC)
Hi! I love this and want to see this concept succeed. I currently have four images that are still hanging around in my queue: three have no tags suggested and one that has an incorrect tag ("court" and it's not a court-related subject). No matter how many times I skip, they're still there. Thanks! Missvain (talk) 05:25, 11 February 2020 (UTC)
Thanks for leaving the feedback about wanting to be able to dismiss images that have been skipped, others are mentioning it above. It's useful to be able to take to the development team, much appreciated. Keegan (WMF) (talk) 22:15, 11 February 2020 (UTC)
Hello. The "tags" now appear in the structured data of the images. However, this seems "doubled up", because the qualifier still has to be added by hand. Missing "tags" can easily be added there. --Georgfotoart (talk) 10:20, 12 February 2020 (UTC)
I love to help tagging, but I don't want to flood my watchlist with all the tagged pictures. Is there a way to exclude them from the watchlist? --Zinnmann (talk) 11:20, 11 February 2020 (UTC)
Found the problem: Under preferences/watchlist I have had "Add pages I create and files I upload to my watchlist" checked. --Zinnmann (talk) 15:42, 11 February 2020 (UTC)
The “tags” are actually on the Structured Data page, but only if they were previously empty. Since some day edits have been undone (due to inaccuracy), I will enter the Structured Data by hand as before. Adjusting the wrong tags does as much work. Best regards --Georgfotoart (talk) 15:56, 11 February 2020 (UTC)
I guess this mis-behaviour is related to Task 216369. I wrote this task with the title "Adding a caption add file unintentionally to watchlist" but it is much wider, see the comments in the task. Raymond16:21, 11 February 2020 (UTC)
Many edits
I tried this out last night. I notice that every tag/claim added with this tool results in an edit - for example, I confirmed nine suggested tags for File:Euston station MMB 99 90019.jpg, which resulted in nine edits to that file from me. In my opinion, that's cluttering the logs. On the other hand I can, for example, add several categories to an image with a single edit. Wouldn't it be possible to implement this tagging in a tidier way, too? Gestumblindi (talk) 20:56, 12 February 2020 (UTC)
Generic or specific tags?
This issue was raised before, but I would like to know where exactly we stand now on this: Should we accept tags that are true, but very generic? People seem to have differing opinions on this. For example, when I tried out the tool last night, I accepted some tags for File:MosOblast 05-2012 Marfino Estate 13.jpg: "arch", "body of water", "arch bridge", and "water". All these are true, I think. There is an arch bridge in the image, there is water in the image. A.Savin reverted this. This doesn't bother me - but I would like to know how exactly we are supposed to use the tool. If generic tags aren't welcome, then we can forget it in its current form, as most suggestions are of that generic type. Gestumblindi (talk) 21:34, 12 February 2020 (UTC)
@Keegan (WMF): Well, I'm not going to tag any additional images using the tool until this has been clarified and we have a meaningful guideline. Some of the suggestions are really extremely broad and generic. One of the suggestions for my recent upload File:Wegweiser in Breitenbach SO Februar 2020.jpg is, for example, "human settlement", and this is 100 % true, of course - it is a village, a municipality in the district of Thierstein in the canton of Solothurn in Switzerland, and the name of it is Breitenbach. Although Breitenbach is in the file name and in the description and in the categories of this image, the tool is not able to suggest the obviously most fitting Breitenbach (Q66672) (maybe other places of this name, a human could then select the correct one, though the category system already has the correct Breitenbach, so it should be possible) but only those super-broad, super-generic "tags". Shall we now tag all "human settlements" from Kammersrohr to Cairo, from Sydney to Nuugaatsiaq that way? I'm not sure how useful this kind of "tagging" is, especially given, as others have mentioned, that we actually would like to add the most precise "depicts" statements we're able to. - Also, the tool gives only the names of the suggested items, but no link to Wikidata, so we can't easily check before publishing whether the suggested tag is really fitting or maybe a wrong tag of the same name. Gestumblindi (talk) 21:37, 19 February 2020 (UTC)
Most tags way too vague, irrelevant or even detrimental -- no tags and just a text box/search would be of more use
Even for images which have clear categories that are tied directly to Wikidata items already, it seems to miss them and instead offer tags from left field. For example, File:Juvenile Nubian ibex (40348).jpg doesn't give me an option for Capra nubiana or Nubian ibex, but does somehow offer two completely different species not mentioned anywhere on the page (Barbary Sheep and Oreamnos americanus) in addition to e.g. "rock" and "wildlife" (not wrong but seems like we'd be doing a disservice to anyone relying on this data to select such items -- or is there an argument to err on the side of tagging absolutely everything like people do on, say, Flickr or imdb). Is it not using the text that's already on the page in favor of [pattern recognition?] (sorry, leaving feedback before doing a deep dive into the mechanics here because of limited time). — Rhododendritestalk | 00:51, 13 February 2020 (UTC)
The suggested tags are too far off. Of the nine of my pictures that got suggested tags only four pictures got relevant suggestions. All nine pictures was already tagged with highly specific tags by me (the correct mammal species or even more specific tags), so the relevant suggestions was too general. In only one case a suggested tag was spot on (i.e. the correct mammal species), but that tag was already present. In another case the picture was tagged with the correct mammal species, but the suggestions was other related species. / Achird (talk) 02:36, 13 February 2020 (UTC)
What am I supposed to do with the pictures with just irrelevant suggestions? How do I get rid of the irrelevant suggestions? Now the only thing I can do is to accept an incorrect tag to get rid of the suggestion, so in it's present state this computer-aided tagging is useless. / Achird (talk) 00:38, 15 February 2020 (UTC)
Error during manual tagging
Some images depict clearly identifiable persons or objects. If that's the case I open the image in a seperate tab and try to enter specific information in the structured data tab. Sometimes that works, sometimes not. In the latter case the following error mesage appears: Invalid value "" for integer parameter "baserevid". This happens e.g. when I try to tag File:FIBD2020DorisonDelep 03.jpg (depicting fr:Xavier Dorison) with Q3570662. Is this is known bug? --Zinnmann (talk) 14:19, 13 February 2020 (UTC)
1. Tagging tool needs to take information from the file names to create tags.
Many file names are specific and accurate enough for the tool "read" the files name to create suggested tags for human approval. Can the tool be coded to read file names, perhaps using OCR, before presentation to the user to select suggested tag(s)?
2. The tool repeatedly displays "skip"(ed) files.
3. The tool shows files with no tags.
4. The tool does not allow manually applied tags by user. Tagging tool should allow user to manually apply tags. Alternatively, the tagging tool could have drop-down menus for general tags with more specific related tags or more related variations of tags.
5. Tagging tool should have and option to have many more choices of tags. Example: Instead of just a few suggested tags presented on the first view of that page, with a mouse click allow an expanded list of tags with a "maximum limit set by user" or just multiple option of numeric quantity of tags to be displayed- such as 10/ 20 / 50 /100. This would allow users to better find accurate tags, instead of easy but inaccurate tags or no applicable tags at all.
6. Tagging tool should allow user to note or identify tags that are extremely inaccurate. Perhaps, to help machine learning or does Google do it all themselves before use at the back end of this process?
Thanks to all who have left their comments, concerns, and other observations using the new computer-aided tagging tool. There's a number of posts here re-iterating some feature requests, which is great to see support for, as well as some posts on areas where the tool could be improved to make it more useful overall. I'll be going over this feedback with the development team in the coming days and weeks as the team wraps up other properties work, and we'll see where we are as far as what we can do with the tool to make it better for everyone. Keegan (WMF) (talk) 22:47, 13 February 2020 (UTC)
@Keegan: Thanks for keeping us in the loop. I hope I'm still in time to feed in some thoughts. I very much share Andy Mabbett's concerns that the kind of tags being proposed are qualitatively different from expectations and use of depicts (P180) so far. It doesn't make sense to mix a low-noise resource with a high-noise resource. I've responded at greater length in the thread on Project Chat, but some concrete suggestions for consideration in short:
Don't add any more depicts (P180) statements until the community has had a chance to have a proper discussion as to how it wants P180 to be used.
Consider creating a new, different property, other than P180 (eg "tag" or "keyword" or something like that - the name doesn't matter), and use that for tags, distinct from depicts (P180), leaving P180 to be much more concrete and specific.
Make sure that the admin corps has an 'emergency stop' button, so they can stop the tool immediately and at will for any reason that might become necessary
Make sure that statements created using the statement can be clearly identified and separately retrieved in Wikibase, and (in time) can be distinguished for specific analysis when SPARQL is available -- eg by using the Wikibase reference facility with a specific reference property and value.
Assuming there is a set policy for tagging (for a color portrait of someone, would we want to tag color, portrait, female, Individual's Name, or just the Individual's Name?), a few standard images with tags to select, where you select them and the application gives you feedback (too specific, not correct, etc) would be helpful. I had a couple that were pictures of a text document, and options for text and document tags were there, and it was pretty unclear which would be preferred. The same with a brook, where I believe natural resources, water, creek, and some others were options. If we have a policy, training to help show users how to apply the policy correctly would be useful. If we do not have a policy, we should develop one before mass-tagging images. Kees08 (talk) 00:21, 14 February 2020 (UTC)
Kind of, but many sections above deal with the same general topic and this provides a solution (training). My suggestion is that a tutorial becomes part of the tool. Kees08 (talk) 19:22, 14 February 2020 (UTC)
Gender neutrality?
I have some concerns about gender neutrality of the tagging system. I have recently added several photos of ambassadors taken by the same photographer, the same background etc. Nevertheless, for male ambassadors there are suggested usually such tags as: suit, official, businessman, white collar, whereas for female ambassadors: beuty, hair, blonde, lips. Niegodzisie (talk) 13:04, 15 February 2020 (UTC)
That is a problem of the ML training data used by Google. This is one more point showing that these ML and the API are definitely in a beta stadium and not ready for real production. --GPSLeo (talk) 09:57, 16 February 2020 (UTC)
Add information from categories
Many images are categorized, and often the category is way more precise than the suggested tags from the AI. Maybe we can include wikidata items from the images' categories as well. Additionally if any Wikipedia page links to the file, this would also make a great suggestion. --Geek3 (talk) 22:25, 15 February 2020 (UTC)
The provided tags are often very unspecific. Maybe the user knows even more about the image or can find out with a few clicks. It would be great to add a field for custom tags that were not proposed, where the user can choose from all tags available on wikidata. --Geek3 (talk) 22:27, 15 February 2020 (UTC)
Hello, I am not interested in adding tags to my photos. How do I turn off notifications? They are annoying. --Venzz (talk) 10:23, 18 February 2020 (UTC)
Verkehrsweg - hard to ttranslate, somewhere between "thoroughfare" and "right of way"
Obviously none of them is useful. I can only "skip", but then the tool will come up with the file every single time I add some pictures and are asked to label them. --h-stt!?16:04, 18 February 2020 (UTC)
I have the same problem. In case of Lichen this is no plant, no grass, or moss ... And how can I set a correct tag in this case?
How to block messages from previously correct tagged pictures (once only block unuseful tags for a picture) to avoid that their massages came over and over again? Greetings --Wilhelm Zimmerling PAR (talk) 08:22, 19 February 2020 (UTC)
In that case a structered data entry "Gewöhnliche Gelblechte" (a species of Lichen) exists since upload. Nevertheless the messages come over and over. And I think, an additional entry "Flechte" (or "Lichen") is unwanted in this and such case. In the special case the Lichen grows on bark (of Acer pseudoplatanus), but it makes no sense to tag these terms, for the picture. What can I do ? (I don't want to disable all such messages or ignore them.) Greetings --Wilhelm Zimmerling PAR (talk) 13:43, 20 February 2020 (UTC)
@H-stt: I have that problem too. There is an unsightly solution: save a tag and then delete it from the image in the structured data. As a result, the image is considered edited. Best wishes--Georgfotoart (talk) 10:55, 25 February 2020 (UTC)
I would say there is no such way. We have to multiply our request to WMF developers to add features to this tool such as the posibility to remove it from list, or give an opportunity to manually add tags or retrieve tags from category name. --Juandev (talk) 06:37, 5 March 2020 (UTC)
Add "DISMISS" and "New suggestions" below the suggested tags
Please, add "DISMISS" and "New suggestions" below the suggested tags. "DISMISS" or another term can eliminate a repeated photo display with no useful suggested tags. "New suggestions" or other term can request a new and different set of suggested tags for a photo with no useful tags, if this computer-aided program is able to generate different suggested tags.
Tags should be suggested according to tags that have already been applied to images of the same creator. For example, I am interested in medieval art and architecture, and often use special tags for this, but most of these tags never have been suggested. I would like to see them as suggestion. --Uoaei1 (talk) 12:14, 20 February 2020 (UTC)
One of my image already had two depicts when was uploaded. Today I used «Computer-Aided Tagging» and now two tag duplicate each other — Araneae and spider web. See file(ATTENTION! — naturalistic spider) and history.--Doomtrooper (talk) 09:43, 21 February 2020 (UTC)
Let me begin by saying that I think computer-aided tagging definitely has a place on wikimedia commons and I've been hoping one of the more technically competent member(s) of the community would introduce something like this for several years now. I found this tool however to be disappointing, mainly because of its weak predictive/suggestive capabilities. I don't know if there is a machine learning element to this tool but if there is then it really needs a better training set and additional iterations before it is deployed.
More specifically, I recently uploaded an album of images taken by a famous British diplomat, Percy Sykes. Within a day or two I received 48 alerts, one for each of the uploaded images asking that I review the suggested tags. I did this and found that the suggested tags were so vague and non-descriptive that it almost certainly be better to leave the image with only the tags that I had uploaded previously or, in the case of an image without any category tags, keep a "category tags needed" tag attached to the image and stop there.
Here is just one example, selected at random from the 48 images i uploaded. The image is titled "The City of Kashgar" (see right):
View of Kashgar in 1915
This image has the caption: " View of Kashgar in 1915 " and the description :" View of Kashgar in 1915 "
The image also had the following category tags when it was uploaded:"1915 in Xinjiang" and "Kashgar"
The suggested tags were the following: Monochrome, tree, photography, photograph, stock photography, snapshot, history, monochrome (again!), black-and-white, and picture frame
The suggested tags reveals several shortcomings:
-1)Duplicate tags can appear (in this case monochrome is suggested twice).
-2)Suggested tags can be children of other tags (e.g. snapshot is a sub-category of photograph). Over tagging can easily occur if these are suggested tags and that will create even more problems down the road.
-3)Most critically, the suggested tags don't appear to use any natural language processing of the existing tags, image title, image caption or image description. If they had they could likely have picked up on "Kashgar", "1915", or "1915 in Xinjiang" and immediately suggested things like "1915 photographs" or "History of Kashgar", both of which would have been excellent and valid suggestions.
I suggest whoever developed this immediate fix the duplicate tags and children issues. Also it could use basic natural language processing of the text for each image to suggest relevant image tag suggestions. Please let me know if i can be of further help.
...and.... how do you switch it off? I want an easy to find button please
I load pictures linked to a Wikipedia article with depicts and the suggested tags are "chin". May be of some use for an "alt" description.
Main improvement is that when I click on "suggested additions" there should be a button on the first screen that allows me to switch it off. Please? Victuallers (talk) 14:28, 21 February 2020 (UTC)
CAT tool usage and updates on effect on search/discovery coming soon
Hello everyone. Early next week, we plan to provide some additional context and information on actual CAT tool usage, impact, milestones, and next steps. Thanks. RIsler (WMF) (talk) 01:24, 28 February 2020 (UTC)
Thanks for the feedback. Basically, these requests are in the works to an extent. The ability to see categories is being built, which should work in a similar fashion to your suggestions about showing the file description. The ability to add additional tags of one's own choosing is being implemented as well. I look forward to seeing what you think of the improvements when they're released. Keegan (WMF) (talk) 22:48, 2 March 2020 (UTC)
I'm afraid it's useless as currently built
As T.Shafee says, the idea is great. Shame about the current realization, then, as it suggests useless tags - I was given an image of some fossil parasites on a mainly yellow background, and all I could do was agree to "Yellow" or just "Skip"! Not terribly helpful; if I could have added "Parasite", "Fossil", "Ectoparasite", "Amber" etc then we'd be getting somewhere. These things could be constrained to terms from the image page to reduce silliness. Otherwise, the whole thing should be placed in a rubbish-Skip. I see there are many other users who've had similar experiences. Chiswick Chap (talk) 09:39, 29 February 2020 (UTC)
Double-up of tags
I have recently started applying tags to images I upload. Now, today, I had notifications to tag 18 recent images. I'm going through the suggested tags and adding the ones I believe are suitable. As I have however already tagged these images shortly after my upload, some are there twice now as I did not realise these were doubled-up and it's hard to remember what tag you already placed a few days ago. Have a look at File:Sutton's Farm, Halls Head, February 2020 03.jpg as an example, which now has "barn" and "farm" there twice. Does this matter and, if yes, can the system be made to recognise that a tag is already there? Calistemon (talk) 12:14, 29 February 2020 (UTC)
enabling users to refine suggested tags before submitting could be useful. e.g. open a tree of subitems on long click on a tag suggestion.--Chewzy (talk) 17:50, 29 February 2020 (UTC)
This system is awful. I just reviewed 11 pictures again, and there was not the LEAST suggested keyword actually related to the pictures (whereas I provided all the categories and stuff while uploading the pictures). Moreover, several suggested keywords were even misleading, and people not aware and knowledgeable enough may turn it into a massive misleading machine. This thing is the contrary to what Wikimedia wants to be. FredD (talk) 11:36, 1 March 2020 (UTC)
Sending a notification on every uploaded file is too much. I'm suddenly seeing 50+ notifications. Only one is needed to get my attention and then if I go to the tagging tool, all 50+ files are offered up to me. Going to the tagging tool should then reset the need for a notification to say "oh, there's more now to tag". Like a watchlist. Having said all that, I felt the tags were fairly generic and didn't seem to me to be add much value. Also you have to tag the files individually. This is where tools like Cat-a-Lot and AWB should be used. Often you are uploading a set of related photos so we need to extend the existing "power tools" to allow these Wikidata tags to be edited with these tools. This so-called structured data needs to be part of the File info where it can be edited with normal tools not by introducing a set of new inefficient tools as is the current situation. Having experimented, I will be opting-out as I think the fundamental approach (building new tools rather than extending existing tools) is the wrong strategy. We worry about the declining editor numbers so I think we need to think in terms of improving power tools rather than introducing extra manual steps that just makes more work for the existing community for what seem to be low-value tags. Can I opt out of the "depicts" step on uploads (especially as we have this tagging now to do the job if people want it). I think the folk who will do this tagging will probably be those who are already be doing an OK job with file naming, description, categories etc. What I think we need more is tools that will generate suggestions for the helpful file names, descriptions, categories etc for the uploads that lack them rather than add these tags. It is unclear to me how Commons and Commoners benefit from these tags and their tools. Kerry Raymond (talk) 13:52, 2 March 2020 (UTC)
Thanks for your input. The notification issue appears to be a regression/bug in the notification system and we're investigating. RIsler (WMF) (talk) 22:54, 4 March 2020 (UTC)
Same for me, I have several in my list and they keep being proposed first whenever accessing the tool. So, I have more and more to skip every day. Romainbehar (talk) 08:37, 7 March 2020 (UTC)
Needs a "none of these apply" button to remove an image from the queue. (I'm getting two suggested tags of "road surface" and "jeans" for File:Gunter_Demnig_installs_Stolpersteine_in_Amsterdam,_October_3,_2018.jpg, a man in a denim shirt on a sidewalk: if I press "skip", it just comes around again.) I'd have expected that going to the image and tagging it manually would have removed it from the queue, but apparently not.
Failing that, those kinds of "stuck" images should be at the end of the queue when I click the Special:SuggestedTags link. It's currently the first image I see every time, and I have to skip it to see the picture you're actually alerting me to.
I don't know if there's a policy definition of the difference between an alert and a notification, but these feel more like notifications ("some optional processing is ready for you to take a look at") than alerts ("you should take a look at this right away").
A FAQ at the top of this talk page, or a section on the project page, saying how to opt back out would be useful. I had to read through a lot of the above comments to find out. --Lord Belbury (talk) 16:10, 5 March 2020 (UTC)
Thanks for your points, particularly the last one. I'll look into getting an opt-out notice more prominently someplace appropriate. Keegan (WMF) (talk) 20:13, 5 March 2020 (UTC)
The suggested tag (computer-aided tagging) on a given image that has been uploaded to Wikimedia-Commons is not always accurate. The app (as it stands) does not allow me to write the correct tag. It only allows me to confirm the selected tag, or to skip it. When I skip it, the image returns and repeats the old process, viz., requesting me to select the correct tag, but not allowing me to write the correct tag whenever the correct tag is not available on the selected options.Davidbena (talk) 18:40, 8 March 2020 (UTC)
How are chosen the images for review?
I just enabled this promising tool and clicked on Images to review, then personal uploads. I'm not convinced at all by the first picture presented to me: https://commons.wikimedia.org/wiki/File:F-GRHF_landing_at_Toulouse_Blagnac.JPG I find this image perfectly categorized: we know the exact aircraft type (A319), registration (F-GRHF), airline (Air France), and at what airport it's landing (Toulouse-Blagnac) and when (April 2013). Yet the suggested tags are so vague I can't see the benefit to add any of them (I get "aircraft", "surface vehicle", "flight", "airliner", "aircraft travel", "engineering and spatial technology", "vehicle", "airline", "aviation", "plane"). Are the Commons categories not used at all to suggest tags? That's where 90% of the accurate information is. vip (talk) 00:10, 9 March 2020 (UTC)
Hello vip, and thanks for your question. We have 2 classes of images: the "Popular" tab contains images that have been assessed by Commonists as featured, valued, or quality. We also allow users to opt-in and have their own uploads analyzed. Categories are currently not used as part of the algorithm, mainly because text matching to Wikidata entities is tricky and prone to errors. However, we'll soon be adding two features that will utilize the info Commons already has in categories.
We'll list up to 3 of the image's existing categories underneath the file name
Users will have the ability to add their own values within the Suggested Tags interface
However, we do also see value in the less specific statements. While searches for very specific terms can work on Commons, for years we've gotten feedback that search on Commons returns poor results and the category page structure is too hard to navigate, especially for students. Current search on Commons is okay if you know you're looking for an Airbus A319 specifically (although this still has problems), but most internet users are accustomed to Google-style searches like "planes" or "airliner" and then drilling down into specifics from there. The current system on Commons leaves valuable applicable/relevant images hidden and often leaves searchers unsatisfied. We'd very much like to make Commons better at this, and we've seen encouraging results from the Suggested Tag system far. We'll continue to refine it as we gather more data and feedback. RIsler (WMF) (talk) 18:14, 9 March 2020 (UTC)
Two of my pictures have no suggested tags ... please correct!
File:Kameldornbaum Sossusvlei.jpgFile:Hippos am Sambesi.jpg
Two of my pictures have no suggested tags ... please correct! These two images continue to show up almost every day but I cannot change any thing because there is no suggestion of any tag. Regards, Hp. --Hp.Baumeler (talk) 08:40, 11 March 2020 (UTC)
Yes, there's no way of getting rid of an image from the list other than by clicking on a tag, and if
@Hp.Baumeler: It appears you've discovered a new bug. The team is looking into it and have discovered the source of the problem. You can follow the progress here: https://phabricator.wikimedia.org/T247460 We will also soon be adding the ability to remove items from the queue without needing to confirm tags, as well as the ability to add custom tags. RIsler (WMF) (talk) 21:42, 11 March 2020 (UTC)
I mentioned above that there's no visible way to unsubscribe. How may users do that, please? It should be in the instructions, preferably actually on each "Suggested tags" page so users don't have to chase around looking for it. Chiswick Chap (talk) 09:59, 17 March 2020 (UTC)
Thank you very much. At least it's now documented here on this page. It should really be a bit more prominent, like on the "Suggested tags" page that the bot generates also. Chiswick Chap (talk) 15:20, 17 March 2020 (UTC)
My feedback
Still showing images with poor tag offerings. I would like
I wonder, what is the quality of those services. Recently we had a Wikimedia Photographers meetup and we were also discussing the use of artificial recognition, but we came to the conclusion it is too early to use it, because the quality of output is bad - i.e. the machine shows more objects with very close probabilities, and for many objects there are not created models yet. So we have concluded, we would appreciate software, which would provide Wikidata tags retrieved from the radius around a photographer (or in the direction of photography taken) to the users (this is described here in the Czech language).
By the way, how much it will cost WMF to use this service. I would say the number of files from Commons, will exceed the limit of free use. Juandev (talk) 05:50, 18 September 2019 (UTC)
The agreement with Google is still in the works, I have no details about it yet.
As to the quality, the short answer is that it varies. For what the development team is looking at implementing here, simple depicts of concrete things, the technology is working pretty well. If you're looking for pictures of a bicycle, the software is pretty good at finding the fixed concept of a bicycle. Where the technology is arguably catching up is in abstracts; while it might be easy for the computer to see and identify a bicycle, it is harder for it to identify the abstract act of cycling. Community policy around depicts will help people know what to use with the suggested tags. Keegan (WMF) (talk) 16:21, 19 September 2019 (UTC)
More noise than precision
Do we need 1 million bicycle identified in pictures? I did a small try and I feel suggestions will just be adding "noise" not adding precision. - Salgo60 (talk) 00:36, 7 February 2020 (UTC)
@Salgo60: Yes, we do need that. If we have a million images with a bicycle on it, we should have it in structured data as well. And just as with categories, there's always the possibility of making an existing statement more specific if needed. --MB-one (talk) 13:27, 12 February 2020 (UTC)
Yes, 1 million pictures of bicycles would be wonderful. That would allow all kinds of further analysis and data mashups. Thanks for asking :-) Syced (talk) 08:52, 21 March 2020 (UTC)
Upcoming improvement to computer-aided tagging
I've said before, but I"ll say again, thanks for the feedback on using the tool. It's helped the team determine what they should work on as next steps. So with that, there are a few things coming up for the team to do here:
Bug fix: permanently dismiss the invitation to tag images from the UploadWizard when a user closes out the notice. This is intended behavior that is broken, the notice currently reappears each time in the UW. There's a patch up for this.
Add own depicts tags from the computer aided tagging screen. By far the most common request and complaint was the inability to add one's own depicts statements as an all-in-one workflow.
Show categories. This feature request will help determine the accuracy of tags to categories already sorted for the image. Users will be able to see the categories and make judgements based on that additional information.
See more information for a depicts tag. Mainly affecting homynyms, users need a way to disambiguate tags with similar or identical names but different concepts, to assure accuracy.
Remove new file versions from personal uploads. Another common issue was that uploading a new version of a file put the file in "personal uploads" and were then brought up for tagging. This behavior will cease.
The team plans on starting to work on some of this soon, the bug-fix patch will be landing first for sure within the next week or so and development will go from there. I'll keep everyone updated as the new features progress. Keegan (WMF) (talk) 18:32, 21 February 2020 (UTC)
@Keegan (WMF): What about the very strong feedback, both here and at Village Pump, that the tool should be feeding a different property, distinct from depicts (P180) ?
There's not really much of a response to give. The tool is built to aid users in tagging depicts statements. "Feeding a different property" as you suggest defeats the purpose of the tool, and from what I can tell other users are not strongly backing your suggestion on the VP or this page. If there's a different conversation going on around your proposal where users are showing such strong support for an alternative property that I haven't seen, I'm happy to share that with the team if it will help change minds and show community support here. Keegan (WMF) (talk) 22:22, 24 February 2020 (UTC)
To be clear, it's not that your idea is good or bad. It's that it fundamentally re-scopes the design and purpose of the tool and would necessitate a lot of engineering time to make it happen, and that time isn't currently allocated as it's going to towards bug-fixing and implementing the other new features. Keegan (WMF) (talk) 01:13, 25 February 2020 (UTC)
At this time (till the tool will work better) it seems to be the best way, I fill in structured date during upload and ignore all proposed tags for my with structured data uploaded files. And the tool should better switched in a beta-phase modus. Greetings --Wilhelm Zimmerling PAR (talk) 07:48, 25 February 2020 (UTC)
No; the problem - which you refused to didcuss when it was raised on the Village Pump - is that the tool fundamentally re-scopes the design and purpose of the 'depicts' statement. You have been asked for evidence that there is consensus for this change, and have offered absolutely none. There has been no support for such a change at VP, on this page, nor, apparently, anywhere else. Turn it off, until you can show consensus. Andy Mabbett (Pigsonthewing); Talk to Andy; Andy's edits10:28, 25 February 2020 (UTC)
The quality of the suggested tags are so low so please dont add them to the same "bucket" as depicts (P180) and also have a qualifier so we can easy track and delete them... - Salgo60 (talk) 21:43, 27 February 2020 (UTC)
This tool is very much needed to help users fill more images' depicts (P180) field. While it still has a lot to improve, it is already rather useful at that (just click "Skip" if no good suggestions are shown). Setting a different property would be counter-productive. Syced (talk) 09:13, 21 March 2020 (UTC)
We'll be adding some items to the system's blacklist to account for some of the stranger suggestions coming from the machine vision API. This will be an ongoing refinement process
The feature changes mentioned above are in progress. We'll probably deploy listing categories for the image first, followed by the ability to add your own custom depicts statements.
The bug that prevented permanent dismissal of the opt-in notice on UploadWizard should be fixed
What is the way to request an item to be added to the blacklist? If there is none yet I think a simple sub page here would be the best way that everyone can easily do such requests and the development team can and should mark them as resolved is they got added. A bit similar to Commons talk:Abuse filter --GPSLeo (talk) 23:38, 4 March 2020 (UTC)
Hello, GPSLeo. That's a good idea. Currently, we monitor suggestions from the system side and add more to the blacklist as a more diverse range of people use the tool and establish usage patterns. Also, changes to the machine vision API on the Google side can affect things and we monitor that. We're working on a couple of different long-term options for community work on the blacklist and we're also considering something like your suggestion as a solution for the near term. We'll put something together soon. Thanks! RIsler (WMF) (talk) 00:21, 5 March 2020 (UTC)
No stats on the number of edits that have been reverted or otherwise overridden? No assay by a human of a random sample of the content created to see if it is "good"? - Jmabel ! talk06:56, 5 March 2020 (UTC)
Hello Jmabel. We're still doing some systems checking to see if the code that checks for these reverts is working properly and accounting for everything. At the moment the system has tagged just over 100 reverts in the past 30 days. As for human assessments, we're constantly looking at both what gets suggested and what people are picking. That's why we're updating the blacklist to account for both user behavior and also inaccurate suggestions coming from the Google API. As an example: as Chinese speaking Commonists have begun using the tool we've noticed a slight increase in the frequency of some users choosing depicts:occurrence on images depicting stage shows, festivals, celebrations, etc. After looking at the Wikidata item in question, we discovered the Chinese label on the Q item is '事件', which translates closer to 'event'. There is an actual, different 'event' Q item on Wikidata that would be more accurate, but the machine vision API doesn't currently map to it. To avoid the confusion entirely in the future we'll be blacklisting the 'occurrence' item from the suggestions and bringing the issue up to Google. RIsler (WMF) (talk) 21:30, 5 March 2020 (UTC)
Still no meaningful response to the issue: "There is a big difference between a 'tag' and what should be a value in a depicts statement". Still on response to the question of how to get the invitation for all autopatrolled uploaders to add "tags" turned off. Still no response to the question about how to have the "tagging" tool returned to beta status. Still no answer to the request for evidence of community consensus for this tool to be used, or for tags to be added in this manner. Andy Mabbett (Pigsonthewing); Talk to Andy; Andy's edits15:59, 5 March 2020 (UTC)
Idea: When I select a class, propose its subclasses/instances
On a picture of an airplane, the tool would typically suggest airplane (Q197), which is better than no depiction, but a more specific item would be better.
Idea: When I click "airplane", the tool should open a kind of popup or something that shows me all subclasses of airplane (with a searchbox if there are too many):
API to retrieve "depicts" suggestions from Commons app?
The Commons mobile app allows Wikimedians to select "depicts" items before uploading a picture (using a search dialog).
It would be great if we could get suggestions from your project. Our app would need suggestions like 20 seconds after the picture is available (but not yet uploaded).
In other words, ideally we would send a request containing picture data (probably at reduced size), and as a response we would receive a list of Wikidata items.
Would such an API be thinkable? Or is there any technical or contractual subtlety that makes it totally impossible?
@Hgrobe and Georgfotoart: With few exceptions, categories should be plural.
"Suggested tags" isn't about categories. It is about the structured data "depicts" property. I don't like it, and am largely ignoring it myself, but if you are going to use it, do please try to learn about it first. - Jmabel ! talk19:39, 22 March 2020 (UTC)
Just saw this tool for the first time. It's awful. It suggests crap, and non-experienced users are adding depict statements that don't make any sense. Please consider taking it back. Scann (talk) 01:58, 24 March 2020 (UTC)
Need an "enter" on confirming adding a tag
Right now, when I select tags and then click "Publish", in the confirmation window I don't have a quick keystroke that can confirm -- especially when working with a touchpad on my laptop, this adds way more time than is needed for that action. Sadads (talk) 13:30, 24 March 2020 (UTC)
I don't wanna see any notifications about this disturbing tool
The suggested tags page desperately needs instructions
What do I select? Everything that could conceivably apply? Every tag remotely about the subject, but not about the background? Only the narrowest possible tag? I have a picture of an insect, and it's giving me a range like insect, [common insect name], [genus], [species], [family], pollinator, invertibrate, etc. If these were categories, only the species would make sense... — Rhododendritestalk | 01:36, 10 March 2020 (UTC)
It just gives a list of proposed tags, which you can select from. With no way to add others, at least at this step. Jidanni (talk) 01:39, 29 March 2020 (UTC)
Not only this is a useless tool, now it's also CONSTANTLY sending me messages saying "your uploads are ready for review" (even when I've added structured data to the images ALREADY), but I also can't find a way to SHUT IT UP. This is VERY ANNOYING. Pinging @Keegan (WMF): Scann (talk) 17:55, 30 March 2020 (UTC)
I agree with this opinion. This tool is invasive and unnecessary. It offers inappropriate keywords that it would be tedious to correct each time. Zythème (talk) 20:10, 30 March 2020 (UTC)
@Keegan (WMF): This should be offered in a simple way. It wasn't obvious where to shut it up (especially after it send me like 10 notifications of the same image that I've already skipped) and users should have an easy way to opt-out from the tool. Right now it isn't. Scann (talk) 19:09, 1 April 2020 (UTC)
@Keegan (WMF): Not to mention that I also find it utterly problematic that users have to opt-out from the tool. I didn't ask for this tool to appear in my preferences, it has been added to my preferences without my consent, and this is especially problematic because it's under the category "Notifications". I believe this violates the current privacy policy because this is not covered by any of the current uses allowed in the privacy policy that the WMF has. This is an invasive tool created by paid developers of the WMF, so the privacy standard should be applied here more than anywhere else. The privacy policy applies to the platform, so the fact that you're sending me notifications through the platform could potentially be considered a violation of the privacy policy. This should be an opt-in tool, I shouldn't have to be figuring out how to opt-out from the tool. These are the things that people can sign up for and opt-out inside the "Notifications" section:
Talk page message
Thanks
Structured Discussion
Suggested tags for review -- the obnoxious tool that has been developed by WMF
Mention
Failed mention
Successful mention
Page link
Connection with Wikidata
Failed login attempts
Login from an unfamiliar device
User rights change
Edit revert
Email from other user
Edit milestone
All the rest of the things fall under a general category of "communications" that are in principle related to my behavior inside the platform and that I should indeed be notified about (for example, if someone is trying to access my account from an unfamiliar device). I'm failing to see HOW the deployment of a non-essential tool is something that has been added to "notifications" when this is clearly not a communication tool, is a tool to do something inside the platform, more like a gadget than anything else. Gadgets are for the most part things I have to sign for, and this is what this tool should be: a gadget that you sign in for, especially if it's going to be handling notifications this way. This tool is basically spamming users. Scann (talk) 19:36, 1 April 2020 (UTC)
@Scann: the computer-aided tagging tool is opt-in, and notifications are sent after a user has opted-in. You can only opt-in through the UploadWizard or preferences; no-one is having this tool forced upon them without consent. I meant to add a notice to this page about how to opt-out once opt'd-in, as I agree that's not clear in the tool's design. I've put one notice up now, and I'll pass along concerns about how to opt-out from the tool from inside of it. Keegan (WMF) (talk) 21:47, 1 April 2020 (UTC)
I wonder if there's a bug in the notifications that's sending them if others are using the tool on your uploads. I'll check it out. Keegan (WMF) (talk) 22:25, 1 April 2020 (UTC)
Requests/suggestions can be made on the talk page using whatever kind of process the community would like. The team will patch in new additions as they come up. Keegan (WMF) (talk) 17:38, 1 April 2020 (UTC)
It would be very helpful to be able to do your own tagging.
I support Sadads intervention. I don't understand why media that have structured data are proposed to be tagged again. Even if so, it would be helpful to display the already existing tags from structured data in order to avoid doubles. Greetings -- Hans G. Oberlack (talk) 07:29, 10 April 2020 (UTC)
Limitation to specific categories
Hello. Is there a possibility that the tool only sends proposals to add tags limited to a particular domain or category? I specialize in horses, I would like to have only suggestions for horse pictures if possible. At the moment 95% of the photos I receive relate to areas in which I am not proficient. --Tsaag Valren (talk) 15:46, 5 April 2020 (UTC)
I'm not sure that is technically possible at this time. However, the ability to see categories and add own depicts statements in the tool is coming soon, which may serve a similar function without the automation. Keegan (WMF) (talk) 17:26, 7 April 2020 (UTC)
No suitable tag. The suitable tag would be Úzká, a street in Škvorec
Hey, I've tested CAT with new features and these are my thoughts:
why there is not a possibility to go back if you accidentally skip the image?
allow to search for a string that is contained. i.e. when I don't know the exact label of the item, now I have to go to Wikidata and find it from there. Why? Why so much loss of time? Why is it not possible just directly from this tool. For example, I was not able to locate "Nová hala pražského hlavního nádraží", I was searching "praha hlavní nádraží". The clue here is that Czech is a very flexive language and the first is not infinitive. By string search does not know that "pražského hlavního nádraží" = "praha hlavní nádraží".
too many steps to add the tag:
clic on add tag
click into the search field
type it in
click on add
click on publish
click on OK - could we avoid some of these steps to make things faster?
to see the categories, it's a great enhancement. But at this moment I would love to have an option (you can just switch it on) to fill also other statements than depicts
it would be also nice to be able to add some qualifiers just from this tool. the motivation is to limit time with file manipulation on minimum
How should I add my own tag if it is not on the list? (this is related to point 2)
I am not sure, maybe it would be useful to have to Publish on the right side because when you are adding new tags it's on the right side and your cursor ends on the right side. That would save time.
Could you negotiate with Google, they will train their system? Or could we develop our own technology and train it? Computer tags are still in many cases to meaningful than those we add manually.--Juandev (talk) 10:48, 10 April 2020 (UTC)
Finding a lot of sexist and racist suggestions
Because the tool is not filtering out abstract concepts, they are being applied in sexist and/or racist ways:
Really, we need to have access to a blacklist for this stuff that is transparent for the community to respond to....Sadads (talk) 17:37, 30 March 2020 (UTC)
+1 Yes, this sollution looks like WMC users has are robots and has inifinite amount of time to dedicate it to commons. --Juandev (talk) 08:00, 12 April 2020 (UTC)
Hiding in upload campaign
Is it possible to hide the notice about the tool(not structured data in general) in an upload campaign? --GPSLeo (talk) 15:26, 11 April 2020 (UTC)
Feels its not tested when you cant skip, feels like more and more people feels this experiment should end... - Salgo60 (talk) 05:34, 30 March 2020 (UTC)
Ok, now I'm stuck in a loop. the system says "You have 16 personal uploads for review.", but it only shows me the same three photos over and over, each one with a single suggestion that is not something depicted in the photo, so I only have the choice to 'Skip". I've tried to just hit 'Skip' over an over and occasionally that works, and it lets me edit a different photo but most of the time I end up back stuck in a loop with just those three. I've also tried adding other tags to those photos but that doesn't get me out of the loop either. Hmmm.Ken Heaton (talk) 19:05, 12 April 2020 (UTC)
Show description and P18 image when hovering a tag
Android Commons app, with thumbnails and descriptions. Which "bow" is depicted by your picture?
Tag names are often ambiguous.
Each tag is a Wikidata item, so it most often has a description, and often an image that illustrates it (P18 property).
Both should be shown if available, when hovering the tag. The Android Commons app does it, see screenshot.
Yeah, let's add something, but not to add to much staff to make it chaotic. Yeah, I was wondering what the tag "Day after tomorrow" means and why I see it with every fifth image. --Juandev (talk) 08:06, 12 April 2020 (UTC)
Feedback on computer-aided tagging: cursor, skipping
Hi, thanks for improving the tagging. Since adding my own tags has been enabled, the activity is much more enjoyable :-) There are just a few things I'd appreciate making a bit more user-friendly.
On choosing "Personal uploads" and "Add tag", one normally starts typing the tag words because s/he is used to the cursor being moved automatically in the box (e.g. from adding Commons categories), for example. Here, however, you have to click the box with your mouse first – and have to keep thinking of it on adding each next tag, which always makes me write a lot of words without them appearing in the box as the cursor is not placed there yet. Could the feature be added, please?
It is impossible to prevent photos from reappearing among the Personal Uploads although they have already got a lot of tags in other ways (by adding directly on the file page, for example; sample photo still among my Personal uploads) and/or they have been skipped many times. Yes, you can think of yet another tag to add, of course, but how about showing a button saying "Do not show this image again"? I was seriously considering disallowing the feature when too many old photos kept appearing over and over (before own tags were allowed)...
The images that I upload have been given tags that I think are relevant. But when these same images come up for review the tags I have given are not necessarily there. It is as if they are dismissed just because they are suggested by one person, even though the uploader should be given more weight to the tag suggestion, because they are familiar with the content. I don't know what the value of this tagging is anyway - it seems a waste of effort. I'm going to dismiss further requests for review until something changes and indicates this is uesful.
Jstuby (talk) 03:55, 27 March 2020 (UTC)
After about 2 weeks, there are some improvements. I am glad that when the images come up for review that tags can be suggested, and I am glad that categories are shown. BUT, for fucks sake, I still see the same images come up for review, even after I have manually added tags to the images, and they don't even have the tag I added a few minutes before. You want this to be ignored forever? Change that shit! Jstuby (talk) 01:34, 15 April 2020 (UTC)
Exclude files with depicts statements
We need to make sure that this tool does not makes people give redundant information to images. If an image has the statement what species is on it we do not want to have the information that there is a plant on it. I think the best way to prevent this would be to exclude files they already have any depicts statement. --GPSLeo (talk) 20:56, 17 September 2019 (UTC)
+1 -- I was going to say something similar, the mockups have a lot of "repeat" content, and we probably need to "infer" what tags are similar, in clusters -- i.e. all "dog" types, etc and suggest that people pick the most important/dominant/deeply accurate ones, otherwise we will probably end up with a lot of complex layers of redundancy that will make depicts less useful for, for example, training new AI, etc. Sadads (talk)
I have a related issue. By the time I receive the notice of the tags for my picture, I already forget that I tagged it. The result is duplicate tags.--Роман Рябенко (talk) 11:40, 18 April 2020 (UTC)
Link suggestions to Wikidata-items
Now it is not possible to click on the suggested statements and open the page of the wikidata-item. That would help to find out witch one is correct, especial if they have the same name in the language then you need these information. --GPSLeo (talk) 23:31, 14 November 2019 (UTC)
I am missing this too. Otherwise, I have to search the meaning of the tags by hand, which is not necessarily accurate, or check their meaning later.--Роман Рябенко (talk) 11:48, 18 April 2020 (UTC)
Remove Image from list
It seem that there is no possibility to remove images from the personal uploads list. That means if no suggestion matches the image will stay there and gets loaded again and again. --GPSLeo (talk) 17:53, 15 November 2019 (UTC)
Yes I agree, we should be able to remove images from the suggestions. I've clicked on tags just to "get rid" of an image. Oaktree b (talk) 14:40, 4 March 2020 (UTC)
I agree. We need an extra "None of the other tags" tag. Every time, I have to skip four images again and again. Ksarasola (talk) 21:24, 1 April 2020 (UTC)
Very annoying that I cannot remove a picture from my personal queue. This makes me feel as if I did not complete something. It keeps asking even when I add a tag by hand by using "+Add tag" at the end of the line of suggested tags. I added the same tag multiple times, but the picture returns into the queue!--Роман Рябенко (talk) 11:56, 18 April 2020 (UTC)
Exact identification
I wonder how a tagger will figure out, when AI will do the right identification of a species. Lets say, there is a butterfly and AI provides you two names of a species. How you as a person, who does not know butterflyes could figure out the AI guess is right? Would links to Wikipedia help? --Juandev (talk) 03:45, 17 November 2019 (UTC)
I'd appreciate to have links to Wikipedia or some other wiki project for this. I had tough time verifying plant species suggestions. By the way, all suggestions happened to be wrong.--Роман Рябенко (talk) 11:59, 18 April 2020 (UTC)
Completely useless & a profound irritation
Your tool correctly tags my species distribution maps as "geographic maps", but such a tag is far less useful than the fact that they are all far more precisely categorised as (for example) "Acacia distribution maps". I now have to find out how to turn this off, as I am getting 100s of useless alerts. As an earlier respondent said: this may be useful for completely uncategorised images, but it is certainly wasting my time badly. MargaretRDonald (talk) 02:26, 15 December 2019 (UTC)
I agree. I just got a few "suggestions" for tags. Apparently the suggestions are meant as "structured data", which I thought should be as specific as possible. Just as an example, I mention my image File:Lamium orvala 2 RF.jpg. This is an image of a plant, balm-leaved archangel (Lamium orvala). The suggested tags are Orobanche (wrong), Salvia sclarea (wrong again), perennial plant (correct, but is that really a useful tag?), Symphytum (wrong again), angiosperms (obviously correct, but really not very specific), flower (not wrong either, but do we need this?), plant (even more general than angiosperm, but correct). I skipped the suggestions and added "Lamium orvala (species of plant)" as a tag. By the way, if I skip the suggestions they will reappear. --Robert Flogaus-Faust (talk) 22:31, 15 December 2019 (UTC)
Hi @RIsler (WMF): , the add tag button works, then the confirm tag window works, then I get a black bar at the bottom saying "Something went wrong and tags could not be published. Please try again later." Doesn't matter how many times I repeat it, the same thing happens. Onceinawhile (talk) 21:06, 20 April 2020 (UTC)
Thanks for the additional info, @Onceinawhile: . We believe there's a system error in rare cases where the Wikidata item for a suggested tag has been deleted and there is no redirect setup. We're working on a fix but it may not be released until next week due to WMF holiday. RIsler (WMF) (talk) 20:18, 21 April 2020 (UTC)
Sometimes tags are set which are already exist. So don't set existing tags and show existing tags with the suggested tags. --XRaytalk15:00, 18 April 2020 (UTC)
Prevent duplicate tags
Is it possible to prevent the tagging of already existing tags. For example, at File:20200312-Red-Rose-YX61FZH.jpg I tagged the image with Alexander Dennis Enviro200 Dart, a tag that was already present. Would it be possible to prevent this from occurring?
@RIsler (WMF). I stop tagging because of this bug. I don't want to do my work twice! Sorry, I don't like to be a test person without being asked. --Hadi (talk) 16:59, 25 April 2020 (UTC)
Too much to continue using
I tried it, but the suggestions just don't match up with most of the photos... It got to be too annoying to skip a photo over and over. Turned off the notifications in my preferences so I don't have to deal with them. Happy to try it again if it improves in the future. Oaktree b (talk) 02:20, 3 April 2020 (UTC)
How to supress further suggestions about an image?
In general, this service is great. However, I keep getting asked about the same images over and over. I can click "Skip", but a few days later, the same image comes back, even though I have nothing more to add. Is there some way to say I never want to be asked about a specific image again? RoySmith (talk) 16:40, 3 May 2020 (UTC)
There is currently not a way to permanently dismiss an image without adding a tag. However, the development team recently added the feature of creating own-tags within the tool; adding a tag should remove the image from the queue. It's not the most elegant solution, but it's what works at the moment. Keegan (WMF) (talk) 19:12, 4 May 2020 (UTC)
Abgeschaltet
Ich habe das Dings jetzt abgeschaltet. Es ist einfach nur nervend, sinnvolle Tag kann man nicht eintragen, weil die irgendwo anders vorher zu definieren sind. Die von irgendwem (eine Maschine?) eingetragenen Tags sind teilweise grob sinnlos. Mal wieder ein ungefragtes Stück Software. --Bahnmoeller (talk) 21:29, 27 April 2020 (UTC)
Frage mich auch, warum die Kategorie bei den Tags nicht berücksichtigt werden? Meisten sind die Fotos schon gut kategorisiert. Das sollte als erstes berücksichtigt werden. --Berlinschneid (talk) 14:30, 10 May 2020 (UTC)
Be as specific as you can
Commons:Depicts#What items to add (linked from the project page) says "Please only add "depicts" statements conservatively at first" and "Be as specific as you can. For example [Messier 87 → Q3841190]". The latter is also what the tool used to (?) say. Does the tool ever suggest items we are recommended to add?
Current search on Commons is okay if you know you're looking for an Airbus A319 specifically [...], but most internet users are accustomed to Google-style searches like "planes" or "airliner" [...]. We'd very much like to make Commons better at this, and we've seen encouraging results from the Suggested Tag system far.
This needs to be sorted out, as repeated in many sections above.
We have a page with instructions (probably reflecting community consensus), and a view from WMF, reflected in the tool, and these very much contradict each other. Is the issue about who owns the depict statement on Commons, the community or WMF?
I'd be glad helping WMF with their goals of "user friendliness", but now the way to do this seems to be in conflict with what the structured data people on Commons are trying to do, so there is more or less an edit war going on, with WMF canvassing users (by offering the tool), to ruin the work by those trying to use the depict statement for more specific items.
Edit warring and canvassing are not the ways the Wikimedia community should make its decisions and choose its workflow. I hope WMF could discuss the issue at hand with the community and be open for searching solutions to the strategic goals, not just trying to defend a tool. We should reach a conclusion on how to work towards our goals; developing, not disrupting, the workflows of the community.
Can both not be true? My interpretation was that because of how our categories work, you can't just go down all the categories from Category:Aircraft, because are you go further down you will start to get irrelevant results. However, because the links between wikidata items are stated, you could, for example, follow all the subclasses of airliner, and iterate this, with much less irrelevant results.
That's certainly the theory, which is why the overtly added tags should be specific.
If the wikibase team don't yet have any technology to exploit that, then this whole thing is rather premature.
I suspect -- though I do not know -- that what search technology that has so far been built based on wikibase does far too much dynamically in response to individual user queries, and caches far too little to avoid frequently repeating the same operations (e.g. presumably the is-a relationships for a given q-item should change only slowly; that should be cacheable, with a way to amend the cache when such relationships change; it should only need to change due to changes in statements relating q-items, not in response to any application of a tag to a particular photo, much less an individual user query). - Jmabel ! talk16:44, 26 April 2020 (UTC)
It is not currently possible to do search traversal on Wikidata, which is why tags should not be specific. This is not about not yet having "any technology to exploit that," it's about the disorganization and incompleteness of Wikidata's data that is the prohibiting factor. Hopefully search traversal should be possible in the future, but lack of it now is not a blocker for this project when the community can work around that through depicts tagging. Keegan (WMF) (talk) 17:33, 27 April 2020 (UTC)
@LPfi: I wrote that usage of depicts should probably be conservative at first, with the primary goal of not having an explosion of depicts flooding recent changes when the software was released. There was never discussion or community consensus about it, my words were taken from the Village Pump and put into a page. The advice was followed, there was no flood, but the at first seems to have gotten lost. If depicts statements are always kept as specific as possible, then in effect all that's happened is the category system is rebuilt in SDC. "Be as specific as possible" was never intended to be permanent when I wrote it over a year ago. Since then, whenever I try to bring up loosening the requirements, my own words are pointed out to me as some sort of policy/consensus that never happened. It's pretty confusing for sure. Keegan (WMF) (talk) 17:33, 27 April 2020 (UTC)
Does that mean that Structured Data on Commons is you? Is there consensus or discussion somewhere on how specific a "depict" should be? I'd like to have an uncontested, stable and consistent guideline before starting to contribute – and I'd think before anybody but a core/test group are invited to contribute. Without it I cannot judge whether one should approve "red" as a "depicts" of a sunset or rainbow – yes, there is a red colour depicted, but is it prominent or the statement specific enough? If some users approve it and others reject "sunset" as not being specific enough, I cannot see this being good use of editors' time. --LPfi (talk) 10:59, 29 April 2020 (UTC)
No, there was never a consensus or discussion anywhere about how specific a depicts statement should be. You can read through Commons talk:Depicts and see a lot of discussion about having a discussion, but there's nothing close to what we recognize as consensus-building. Keegan (WMF) (talk) 16:11, 29 April 2020 (UTC)
So: if we can't perform some of the most basic database queries on structured data, including that we cannot trace upwards on relationships such as "instance of" and "subclass of", what exactly is supposed to be the advantage of introducing this? Because it would have been pretty easy to attach something to our category pages to indicate "instance of" and "subclass of" relationships. - Jmabel ! talk14:47, 29 April 2020 (UTC)
The point of this is to make your media discoverable to the end user. You don't trace upwards; using not-narrowly defined depicts statements, users can actual find relevant media using search. If you keep narrowly defining things with categories as your frame of reference for depicts, you locking away your media again in category trees that end users don't use. Keegan (WMF) (talk) 16:20, 29 April 2020 (UTC)
It seems that you are saying that all those claims for Wikibase making things more discoverable reduce to saying "Let's get rid of COM:OVERCAT because we thing that makes things more discoverable." If that's the goal, why introduce this expensive new mechanism to do it, why not just get rid of COM:OVERCAT?
Also, though: isn't this completely opposite to the policy adopted within Wikidata, where I believe that the rules for using instance of (P31) are exactly analogous to COM:OVERCAT?
The big selling point was supposed to be the advantage of structured properties describing relations over Commons' looser folksonomy, where it is harder to identify which category-to-parent relationships are is-a relationships. But if you are just going to throw in the tags all the way up the hierarchy, then that issue becomes moot. - Jmabel ! talk21:57, 29 April 2020 (UTC)
It's completely opposite to the policy adopted on Commons. I've asked WMF staff posting here, and asked several times, to provide evidence of consensus for the model they apparently want us to adopt, and they've continually ignored that request. Andy Mabbett (Pigsonthewing); Talk to Andy; Andy's edits18:09, 18 May 2020 (UTC)
Browsing a specific suggested tag
Could the tool work the other way round too by suggesting pictures for a tag instead of suggesting tags for a picture? For example, if I know what Oxalis tetraphylla look like I could go through all pictures where such tag is suggested to assign it, or to reject the tag, or to skip if unsure. --Роман Рябенко (talk) 06:28, 23 April 2020 (UTC)
Well it doesnt work that way now. Basically the tool doesnt work on all images on Commons, but just on sets of images. So here would be better, you can set a specific category to be tagged. You can already do this with ACDC.--Juandev (talk) 22:44, 10 May 2020 (UTC)
Thanks. I haven't updated it yet because there are/were still elements in design, and I didn't want to give the impression that things are complete when they are not. I plan on modifying the section in the near future as I think work becomes actually complete. Keegan (WMF) (talk) 18:25, 18 May 2020 (UTC)
Question
Should one simply confirm and add abstract tags as suggested (such as "sky", "architecture", "water") or should one also add the name of the concrete object depicted (i.e. name of the building, the river)? The latter is possible, but is it also desired? --Superbass (talk) 19:10, 5 May 2020 (UTC)
@Superbass: Specific tags are absolutely desired. Some of us feel that is far more desired than all of this "sky", "architecture", "water" stuff. - Jmabel ! talk00:33, 6 May 2020 (UTC)
@Pigsonthewing: I have no idea what your "no" means here. The original poster asked "should I do A or B," so "no" makes no sense if that is who you are replying to. Or are you replying to me, in which case could you flesh out your disagreement beyond, "no"? - Jmabel ! talk20:36, 18 May 2020 (UTC)
I missed the OP's second part; my "no" was in reply to their first part. A more complete answer would be "no to the first part; the second is what is desired". Andy Mabbett (Pigsonthewing); Talk to Andy; Andy's edits08:27, 19 May 2020 (UTC)
Suggested tags
Why don't you propose tag according to the title and the categorization ? The proposed tags are "sky", "picture", "geography", etc. Although all interesting tags can be found in the title and in the categories ! Borvan53 (talk) 19:51, 17 May 2020 (UTC)
The software isn't sending any information along to the third-party image analyzer, it's only sending back what it sees in return. This is why we're showing categories and other file information now, to help contextualize the image for Commons. Keegan (WMF) (talk) 18:13, 19 May 2020 (UTC)
real tag property
I think it would be the best to create a real "tag" property like people know them from other image platforms. With these property there would not be a conflict between these tool and the most specific depicts statements. I opened a proposal for this. --GPSLeo (talk) 19:14, 18 May 2020 (UTC)
Thanks for sharing this over here, I've been keeping an eye on it since it opened to see if it gains any traction to share with the development team. Keegan (WMF) (talk) 18:14, 19 May 2020 (UTC)
Good question. Information about the file is not sent to the machine vision provider for analysis, only the file itself. Because of this, the software is unable to use the title. However, this is the reason for introducing showing the title on the tagging screen, in case users find it helpful in confirming a tag or creating their own. Keegan (WMF) (talk) 16:03, 20 May 2020 (UTC)
Duplication prevention does not create error message
I think there is an duplication prevention implemented. I tried to add a statement that already was at the file. That does not got added a second time. That is good. But there is no warning message there is the regular "Tags added to image" message. --GPSLeo (talk) 21:04, 22 November 2019 (UTC)
My experience is different. I have the same statement added to the same picture again and again. Here is an example. Example with three "depicts:pysanka" tags --Роман Рябенко (talk) 12:08, 18 April 2020 (UTC)
I recently reverted several edits from computer-aided tagging. Some of them were duplicate tags. It is nice that the categories are shown. But pre-existing structured data tags should also be shown. If there is a warning (which I have not tested), then this warning does not prevent users from adding duplicate tags. --Robert Flogaus-Faust (talk) 13:20, 23 May 2020 (UTC)
Excluding images with "depicts" statements from computer-aided tagging might be an even much better alternative. This was suggested by User:GPSLeo in September 2019, but apparently this has not been implemented (yet). --Robert Flogaus-Faust (talk) 13:43, 23 May 2020 (UTC)
I confirm that this is an issue. Sometimes the tags are ridiculous and none of them fits the image, but you can't skip, nor can you add your own images or mark the tag as unsuitable. Teemeah (talk) 10:41, 7 April 2020 (UTC)
I'm getting this too, for some photos. In one case it suggests 'plain' and 'suburb' for a photo which is actually an aerial view of a large Electrical Substation located far out in a rural forest area. I hit skip and it keeps returning the same images, with the same suggestions. There needs to be some way to tell the tool that I've rejected that suggestion, so it won't keep returning it.Ken Heaton (talk) 18:31, 12 April 2020 (UTC)
This needs to be fixed urgently. Otherwise it is too tempting for editors just to "get rid" of an image by selecting a tag which is far too generic to be useful. For example many outdoor photos have "sky" as an available tag, but how useful is it to put a sky tag on a photo which just has some blue in the background. AlasdairW (talk) 21:13, 27 May 2020 (UTC)
Is there a possibility for human feedback if we see that a specific not-so-good suggestion is particularly common? For example, in the few pictures I tried this out with so far, there seemed to be a strong tendency to suggest "hill station" for any picture that shows houses in a mountainous environment. However, as we can learn from Wikipedia's article en:Hill station, this is a quite specific term of colonial origin and not suitable as a generic term for mountain settlements around the world. For example, there are no "hill stations" in that sense in Switzerland, but still the tool suggested that tag for some pictures from Switzerland. I find this rather worrisome, as many users of the tool (like me) will not be native English speakers and may think "oh, hill station? I don't know this expression, but it seems plausible enough, let's accept this tag". I think it would be great if the tool not just contained the possibility to not accept a suggestion, but to report it - something like a "report as wrong" button. Gestumblindi (talk) 21:05, 12 February 2020 (UTC)
I'm getting this one too, quite often, as I live in an area with hills and houses, but I'm not taking these photos in Africa, India or in Australia, where this might be appropriate.Ken Heaton (talk) 18:35, 12 April 2020 (UTC)
I am also getting this often, when it should be a rare suggestion. I have had this pictures of coastal villages, and for lowland settlements with some hills in the background. Please only suggest this when it is clear that the buildings are on a hill above 1000m. AlasdairW (talk) 21:21, 27 May 2020 (UTC)
Wide User interface (UI)
When there are no suitable tags, i must click [+Add tag] in the bottom of monitor
then pop-up window in the middle, but without focus, so I must click to it
now I must click to [Publish] on the bottom left corner
and now [confirm] in pop-up on the top of the page
It would be better, if these buttons are in one area and not on whole 27" monitor. And give focus to both pop-ups. JAn Dudík (talk) 09:37, 29 May 2020 (UTC)
It's nice to have an 'add tags' now, but after the final confirming of the selected tags nothing happens, - nada - niente. And if I restart the process, the picture is still untagged. --PtrQs (talk) 07:04, 9 April 2020 (UTC)
constantly having "Suggested tags are ready for review"....
I have these notifications many times a day... and when I click on it to go and reviews tags... there is NOTHING ! how can I disable these notifications ?
The Disable notifications leads to my preference page, where the "Suggested tags for review" is checked... but I cannot unckeck it !! this is invasive and totally useless... on more than 50 notifications, only ONCE did I have tags to review :( --Hsarrazin (talk) 15:58, 12 May 2020 (UTC)
You have got a global account. I have also got a global account. In my case I cannot check or uncheck "Suggested tags for review" under the "Notifications" tab of my local Wikimedia Commons preferences. Everything is grey there, and blocked. If I click on the "User Profile" tab, however, I can click on the button "Set your global preferences". Then, when I go to tab "Notifications", I can check or uncheck the feature. Best regards --Robert Flogaus-Faust (talk) 21:49, 12 May 2020 (UTC)
A page that asked me to confirm suggested category tags for an image I uploaded, wikilinked to this page/topic. Hoping it would clarify the process, I was astonished to find this page even LESS helpful. In short, it's all arcane, mysterious jargon to me -- nothing helpful at all. And I'm a career technical professional, with numerous computer certifications. Maybe this needs some dumbing-down -- at least, say, to the computer technician level? — Preceding unsigned comment added by Penlite (talk • contribs) 12:03, 28 May 2020 (UTC)
I wanted to translate some parts of CAT, but I was not able to locate the messages via ?uselang=qqx, trick. Basically nothing happened. How to find messages for translation? --Juandev (talk) 17:31, 31 May 2020 (UTC)
I am not entirely sure what you are asking here. But this is a talk page of Commons:Structured data/Computer-aided tagging. On this page you should find a link to Special:SuggestedTags. There you can try the tool for computer-aided tagging. However, this should be done with great caution according to the suggestions on Commons:Depicts and you should avoid duplicate, wrong or (at the present time) too generic tagging. Unfortunately, the tool does not show the "depicts" statements that exist already. However, I am most certainly not an expert for structured data or computer-aided tagging. So, may be the experts can give you some advice that I cannot. --Robert Flogaus-Faust (talk) 22:14, 2 June 2020 (UTC)
@Robert Flogaus-Faust: Today has been the first time Wikipedia alerted me regarding suggested tags and, when I went to Special:SuggestedTags, it said, "Images for review," and actually had images for review. Given the number of similar comments I assume that Wikipedia automatically alerts editors that it has suggested tags prepared before it actually has done so, but if one waits long enough it might actually have "suggested tags" suggested. Hyacinth (talk) 08:13, 5 June 2020 (UTC)
Given five separate assertions of this feature's failure to operate I think that the answer to what I am asking is clear, a feature the does what it says instead of a useless broken feature that points out how Wikipedia eats at time like rust. But whatevs, if it works for you you can pretend that it works for me. But can either of us prove it? Hyacinth (talk) 05:45, 10 June 2020 (UTC)
@Koavf: We're looking into this. It appears that for some users, notifications are sent before the last uploaded image has gone through the queue. RIsler (WMF) (talk) 15:58, 18 June 2020 (UTC)
@RIsler (WMF): This happened again. What is an ETA for when this will be fixed? As a WMF person, can you please give me some insight into the organization's understanding of the roll out of structured data on Commons? E.g. would you rate this internally as excellent, good, fair, poor, or a failure? I'm having a hard time getting any understanding on this. —Justin (koavf)❤T☮C☺M☯06:52, 23 June 2020 (UTC)
@Koavf: We suspect we've found the cause of the issue and have posted a patch. The echo notification system seems to not account for some of the content filtering that takes place. We're investigating further for additional cases that may case this problem. RIsler (WMF) (talk) 19:16, 23 June 2020 (UTC)
@RIsler (WMF): Just happened again. Thank you for notifying me, tho. My above questions weren't just hypothetical or rhetorical: I'm genuinely interested in this. Have the WMF published any internal reports on structured data on Commons? —Justin (koavf)❤T☮C☺M☯20:58, 23 June 2020 (UTC)
@Koavf: The patch just went up today and needs to go through code review before it hits the production servers. Once we're sure it's ready, we'll try to get it in place ASAP. Until then, you may encounter a few more notifications. The WMF has sent annual reports to the Sloan Foundation (which funded development via a grant). The Year 1 report is here and the Year 2 report is here. The Year 3 report is not yet public, but over 10 million Commons files now have structured data directly added via the WikibaseMediaInfo extension and API developed during this project. RIsler (WMF) (talk) 23:02, 23 June 2020 (UTC)
This was already asked in in april, but i will push this up as it seems to be important feature. Based on code if I understand correctly it uses generator=unreviewedimagelabels (Api Sandbox, Repository.php, ApiQueryUnreviewedImageLabels.php) for querying images. So limiting to specific category would need (in this order):
Category filter for unreviewedimagelabels API call
method to read the filtering category name from user (url parameter for special:SuggestedTags page without any UI would be fine for a first iteration)
It would be also useful to be able to select images via what links here (ie. using information from pagelinks or templatelinks database tables) --Zache (talk) 05:53, 11 August 2020 (UTC)
I do not understand the logic of why a song would ever be a suggested tag
Hi @Jmabel: . That particular edit seems to have come from the Android app's Suggested Edits, which actually doesn't have any suggestions from Computer Aided Tagging (it just shows an interface asking the user to enter a depicts value manually). I suspect that the Android user intended to say this image depicted summer time (the season), and chose the first Wikidata item that looked "right" in the search dropdown. The Wikidata item in question doesn't have description text to help disambiguate in any language other than Dutch, which is at least part of the problem. RIsler (WMF) (talk) 20:05, 10 August 2020 (UTC)
@RIsler (WMF): Thanks. I'm glad someone at least has a clue what's going on here. Will what you just worked out somehow to into improving something? So far, when "depicts" have been added to my pictures, actively bad edits like this have outnumbered actively good ones. (I count things like "person" or "tree" as neutral.) - Jmabel ! talk22:45, 10 August 2020 (UTC)
Both WMF and WMDE are monitoring usage and making tweaks where there are clear fixes to be made. WMF is trying to facilitate that process by offering a variety of tools (SuggestedTags and manual edit interfaces) and there are community tools as well. Magnus added structured data on Commons support to QuickStatements, and there's WMDE staffer Lukas Werkmeister's bulk editing "AC/DC" tool that he wrote in his free volunteer time. But some of this process will simply take time and a bit of trial and error as more people use structured data and bring with them more perspectives, use cases, and habits. Like any other type of Wiki content, there will be edits that others disagree with, and edits that are objectively incorrect (either through vandalism or honest mistakes). Some of those mistakes, like the one in question, will be partially inherited from data quality issues on Wikidata and WMDE and the Wikidata community continue to address those. Where possible, we'll make changes to the user interfaces to account for potentially confusing data, like we did when we added a disambiguation view to Computer Aided Tagging. RIsler (WMF) (talk) 19:12, 11 August 2020 (UTC)
It uses Google Cloud Vision? But it's nothing like Google!
It could not recognise Taiwanese President Tsai.
There have been many complaints about tags being too vague, so I tested using photos of celebrities. If you google reverse this same photo, google tells you she's Tsai Ing Wen. So, this Google Cloud Vision is probably way worse than the actual thing they use for google image?--RZuo (talk) 21:59, 19 August 2020 (UTC)
Hello @RZuo: . Reverse image search is a separate technology with different methodologies that can certainly be useful, but also prone to frequent errors/deficiencies since it relies on the textual context of the page that hosts the image, and that can often be misleading. The Google Vision API purely looks at the content of the image and does not factor in how that image may be used in other places. Additionally, the Google Vision API provides us with specific IDs for each concept so we can match to Wikidata items and attempt to clarify ambiguous terms, something that reverse image search does not provide. In the specific case of identifying people, the situation gets a little complex. The Google Vision platform does have an option for identifying "celebrity" faces, and the data behind that is probably integrated into their reverse image search in some way, but the system's definition of celebrity is heavily skewed towards certain regions, certain types of "notable" people don't always count, and the external access to the feature requires an additional API cost that would essentially double our API expenses for this system. Even if we could use facial matching features reliably, we probably would not because of privacy concerns (plus, the CAT tool currently is set to ignore most images containing people, your image just slipped through the cracks because of a Wikidata item redirect issue that will soon be fixed). We will continue to explore options for reliable specificity as machine vision tech evolves and more options become available. RIsler (WMF) (talk) 21:23, 20 August 2020 (UTC)
It knows it sees a fancy tower, but cant give the exact answer. Google Image can.
OK... so it's tweaked to not identify individual people...
but it cannot recognise Eiffel Tower either...
This photo might be of some interest, because it's a scan of a 1996 photo uploaded to flickr only on 21 June 2020, so it's probably not replicated elsewhere on the Internet yet. Nevertheless, Google can find lots of similar photos, and give the definitive answer - Eiffel Tower. It's probably not aided by textual context of the page that hosts the image either, because I've tried putting it on an imagehost such that it has a gibberish filename and no accompanying text, Google Image sees it all the same...--RZuo (talk) 16:39, 29 August 2020 (UTC)
Hello @Jmabel: . We consider "Accurate but not perfect" as a step forward on a path to the optimal situation. Like any wiki content, getting it perfect the first time is a bit of a challenge but steps can be taken to get there gradually. Of course, ideally we'd love all structured data contributions to be 100% spot on the first time. For a variety of reasons, that's simply not feasible right now. But the steps that are taken now do help get us to a result that is an improvement. Already, we're seeing that some files that were previously difficult or impossible to find via text matching search are now discoverable because they have structured data, even data that isn't very specific. These data additions that facilitate discovery are often made via one of the suggested edit tools. This is a step towards making user contributions more visible and useful, which of course everyone wants. However, although we're seeing many instances where even more generic contributions make immediate positive changes, in some cases these first steps will turn out to fulfill a function somewhat similar to maintenance categories - temporary additions that can signal that more work is needed and provide a starting point. There are a couple cases where someone has written a "specify bot" of sorts that goes through files with the more generic data and attempts to use other data signals (categories, templates, captions, etc.) to add a more specific structured data statement. We may see more of that type of work in the future, but we also encourage prolific uploaders to add Structured Data to their own files because they're generally the most knowledgeable about their files and the odds of getting it right the first time are higher (although not certain) and we continue to explore ways to improve those odds. RIsler (WMF) (talk) 22:25, 25 August 2020 (UTC)
@Jmabel: Varying degrees of value, but yes. Here's another way to look at it - it's not all that different than descriptions. Commons requires descriptions on upload, but those descriptions aren't always actually very descriptive or specific (often people only fill it out because it's a required part of the upload process and even then they just want to do the bare minimum and move on). Still, despite not always providing the ideal content, it remains a required field after all these years because having something in that field, even if it's not great, is better than nothing (as long it is accurate). More specific and detailed descriptions are of course desired, but community members don't always provide that. What they do provide is still at least somewhat useful, even if it's just as a starting point for someone else to provide more detail. The same is true for categories - people may add general categories for their image and someone else comes by and changes it to a more specific one (this is one of the most common activities on Commons). The same general Commons usage patterns are going to apply to structured data for a while. RIsler (WMF) (talk) 00:56, 26 August 2020 (UTC)
@RIsler (WMF): I believe you are saying that even on images that are very well described by categories, vague depicts are a plus. Assuming I have understood correctly, can you tell me why that is, and what value they bring? - Jmabel ! talk02:25, 26 August 2020 (UTC)
@Jmabel: Much of this comes down to the fundamental benefits of structured data - better support for multiple languages, improved discoverability, and consistency/predictability in data so that other software can use it. On languages - one of the most consistent complaints we get about categories is that they are largely only in English. The multilingual nature of Wikidata helps us reduce this problem. So even if a file may have good category coverage, if I'm not a great English speaker I'll have some difficulties. But the multilingual labels on the structured data items give me a vector to find things via search and other tools. That leads to discoverability - we have looked at the search data to see what terms people are using to attempt to find things on Commons (either on Commons itself or via search engines like Google). Although some users are certainly looking for specific things, many of the searches are for more generic terms (some samples: forest, painting, computer, space). This is usually attributable to people looking for files to use outside of the Wiki projects (blogs, news sites, etc.). We've done user research and user interviews with these "external users" and they consistently say that, while they appreciate that the Commons categories are at least some effort to organize things, the whole system is far from ideal and people generally dislike using it. Depicts statements, even generic ones, help us take a step toward improving search and discoverability for those types of users (and we are currently working on updating Commons search to take advantage of all the new structured data being added from various sources). And finally, there's programmatic use. In addition to human beings, we have a number of software platforms, both community-developed and by tech companies, that are eager to use Commons in better ways but the existing structure makes it very difficult for them. Categories can be a little useful, but many of them are quite specific and require a level of natural language processing that's often prohibitive. Structured data can make that work more straightforward and help expose Commons files for a wide variety of platforms, including external search like Google. RIsler (WMF) (talk) 19:20, 26 August 2020 (UTC)
As for discoverability: couldn't we have accomplished that a lot more cheaply (in every sense) by something focused on getting names in many languages onto Commons categories? All normal text search tools would get the benefit of that. - Jmabel ! talk23:05, 26 August 2020 (UTC)
& It seems to me that that plus something like Wikidata properties to express the relationship between a category and its parents would accomplish most of the programmatic gain. - Jmabel ! talk23:07, 26 August 2020 (UTC)
WMF did explore the option of "structured categories" and found it would still be very costly. Commons has millions of categories, many of which have cultural or language quirks. Translating and structuring all of them and keeping that updated became a task deemed too costly when Wikidata already had systems in place like multilingual labels, descriptions, aliases and an established system for describing data for visual media in the form of paintings. Additionally, even if we could be successful with translation of millions of categories and modeling the parent-child relationship between all of them, there was still the issue of modeling the many attributes of each individual image within the categories. Structured data was still the best option for doing that, so it made more sense to invest in that approach to cover the many use cases involved with one system. RIsler (WMF) (talk) 00:25, 27 August 2020 (UTC)
I'd say the basic idea is that if you search for "red, bike", eventually you will want to find images of red bikes, not just images tagged as "depicts bike, qualifier red". The downside to this is that from the perspective of the datamodel, some concepts will soon become VERY big.. and that basically only statements marked with primary, are useful as descriptors when figuring out what an image contains. I think the basic failure of the computer aided tagging is that it is VERY bad at finding more specific tags that can match. It doesn't drilldown or bubble up. It has not method atm, which allows it to see tag "university" + tag "washington" and conclude: "Maybe i should ask the human about 'University of Washington'....". This cross referencing between children which are part of classes, is what is missing in my opinion right now. —TheDJ (talk • contribs) 12:55, 26 August 2020 (UTC)
@TheDJ: Thanks for you comments. We are currently working on totally revamping the Commons search experience, that includes updating the search to be able to support a single phrase like "red bike" and return both the string matched results and results matching structured data (qualifier support will take time but basic terms are already working). So, in the very near future you'll be able to search for "bike" and get results with matching text in the filename/description/captions but also files with "depicts: bike" that may not have any of that text to match (or it could exist but in another language). In terms of depicts statements, I want to take a moment to clarify that the "university" and "washington" statements in the image Jmabel linked to above were added via the Android app, which doesn't do any computer-aided tagging at all. It simply presents the user with a prompt where they can manually enter depicts statements while viewing the image, its title/filename and/or the description. So those two tags came from a human deciding those were the tags they wanted to use, given the information they had. So, while I agree that there's room for improvement in the computer vision system, we've repeatedly seen that users, when coming up with their own tags, often don't get very specific either. I agree that there's an opportunity to build in some logic so the system can "figure out" other things to suggest, but that kind of logic is quite hard to do reliably and across a wide variety of topics like we have on Commons. With that said, we have started some work along those lines for related search suggestions. If that goes well, hopefully we can build on that and create a supplement to topic identification too. RIsler (WMF) (talk) 20:16, 26 August 2020 (UTC)
@RIsler (WMF): Something we ought to be doing is recording in the structured data where the tags are coming from, so that this is available to SPARQL queries, templates, display etc. This is what the 'references' capability of Wikibase was made for. Similarly it would also be useful to be able to record and access the provenance of tags which have not come from the computer-aided tagging, such as those that may be of professional quality that may have been imported as metadata with image releases from major GLAMs. It is important to be able to distinguish the one from the other, and to have that metadata sourcing available in the structured data, so that it can be looked at routinely and at scale. Jheald (talk) 21:06, 26 August 2020 (UTC)
@Jheald: Thanks for pointing that out. We do record structured data edits from a wide variety of sources, but it's currently done via mediawiki edit tags. Currently that's the preferred way since not all systems use Structured Data yet and we need this info for other MediaWiki systems like reversion tracking, etc. All additions from computer-aided-tagging have an edit tag, and we do differentiate between manual additions and selections of suggestions. Similarly, edits from Android are also tagged. This is also true of community-developed tools - structured data edits from both QuickStatements and the AC/DC tool are tagged. RIsler (WMF) (talk) 21:28, 26 August 2020 (UTC)
@RIsler (WMF): That's interesting. So what would be the best way to access that information from within a SPARQL query? I know various things are possible using the SERVICEwikibase:mwapi call, but it's not a part of the system I am very familiar with. If one has eg identified a set of files and their depicts statement in one's SPARQL query, would it be straightforward using wikibase:mwapi (or some other means) to identify which of those statements had come from the computer-aided tagging? Or, say, to use wikibase:mwapi or some other means to get a sample perhaps of 10,000 computer-aided tags in one's SPARQL query, so one could look at their characteristics? If one wants to access the edit-history tags from within a SPARQL query, to get at what the provenance of the information is, what is the best way to do this efficiently? Jheald (talk) 15:13, 1 September 2020 (UTC)
Hello @Jheald: . You can indeed use mwapi as a way to call mw api endpoints from within SPARQL. But it probably won't be an efficient process, and the current setup for the query service might not support the load since there may have to be one API call per SPARQL result. You might want to try starting with small sample sets and slowly scale up from there to see how things go. RIsler (WMF) (talk) 18:00, 3 September 2020 (UTC)
@RIsler (WMF): Mediawiki Edit tags are nice for sure, but I do think that @Jheald: point still stands: having the origin information recorded additionally as Wikibase reference would be extremely useful − I can see two reasons for it: 1/ now that we have a SPARQL endpoint, we can query for it (even assuming the right API call can be somehow made through `mwapi`, it’s for sure more expensive and less easy to do [querying references is within the range of my SPARQL-fu, querying the API via SPARQL hardly is]) and 2/ it can be displayed in the Wikibase UI, while edit tags cannot. Which, really, has all been discussed already at phab:T253053. Jean-Fred (talk) 12:05, 4 September 2020 (UTC)
@RIsler (WMF): Technically SDoC supports already the storing of the references just like wikidata does.
Try it!) and also the tools like quickstatements supports it example. The information just is not visible in the UI. Is there some reason why computer-aided tagging would not just populate the reference field when the new tag is saved? --Zache (talk) 19:22, 4 September 2020 (UTC)
Hello @Zache: and @Jean-Fred: There are two elements at play here. First, in the past there have been community debates regarding how/when WMF should just add structured data on our own. We currently work on the principle that every content edit must be made by a community member, even if it's a community member's bot or a user just confirming something suggested by a tool we built, the actual edit and the revision has to point to a community account. However, we're more than happy to take on the role of automatically converting metadata to structured data if there's support for it. The issue has been brought up previously in tickets like phab:T245861, and we're keen to do this work if circumstances permit. Alternatively, we can continue with the current situation where community-written bots handle this sort of work (currently there are volunteer bots converting license data in templates to structured data, for example). The second issue here is support for references. For some rather thorny technical/logistical reasons, both WMF and WMDE thought it would be best for Commons not to support references. If there's a great enough demand for it, we'll see what we can do. But we'd prefer a different property to be used for the specific purpose of defining which tool structured data edits came from. RIsler (WMF) (talk) 19:56, 4 September 2020 (UTC)
Sure, A specific property is a good idea. My example was just for showing that the references were technically working and not for defining what the specific property should be. --Zache (talk) 06:07, 5 September 2020 (UTC)
@RIsler (WMF): " We consider "Accurate but not perfect" as a step forward on a path to the optimal situation." It's not clear who "we" is in this context; but the more significant question is "do the Wikimedia Commons community consider what you are doing to be of benefit?". I also note that you have yet to answer the questions I put to you in February (and have reminded you of, more than once, since): "where is consensus shown, for the tool to operate, or to use depicts statements in the manner that it is" and "how can the tool, or the invitation to tag, can be turned off?". Will you answer them now? Andy Mabbett (Pigsonthewing); Talk to Andy; Andy's edits18:56, 3 September 2020 (UTC)
Observations while trying the tool
I was playing around with the suggestions and I got File:Leaning Tower of Niles Illinois-0272.jpg. You can view the suggestions I got on the page information: "architecture, dome, tower, facade, ancient Roman architecture, landmark, building, classical architecture, National Historic Landmark". I was wondering how a manual edit looked so I looked up Leaning Tower of Niles (Q6509808) and added it. Interface says it added it and I look in my contributions. Nothing happened. Turns out the uploader already added it. Why are you showing this picture to me? Just don't offer any images that already have depicts (P180) or if you want to be more advanced: Don't offer any images that already have depicts (P180) -> <some iterm> and <some item> has instance of (P31) and not subclass of (P279) (we already found the most specific thing).
Some other points:
The user interface does seem to have been improved. Editing was quite smooth.
How do I get the suggested tags to work on a specific image (for which I know suggestions exist)?
Any plans to add the ability to add suggestions? I got plenty of those
Any plans to control what suggestions I get? I would love for example to work on all images that have "church building" as suggestion or based on some other search criteria
Hello @Multichill: . Thanks for the input. These are good points and I'll try to address each directly.
When we first started with the tool, we debated whether to exclude images that had any existing P180 statements, but we received user feedback that users still wanted to add data for files that may have had some data but not enough. So the current system allows for files that have exactly 1 depicts statement. However, usage patterns have shifted since we first launched this tool in December and we'll definitely look at tweaking the algorithm.
Thanks for the comments on the user interface. We continue to think of ways to make it even better as we iterate.
Currently, there's no way to use the suggested tags interface for a specific image that you didn't upload. This is partially because most of the current usage is from users editing their own files and its easy to present all of those recent uploads in a separate queue. We'll look into a simple filename search feature that would allow people to see if a specific file has suggestions. In the meantime, as you've discovered, visiting the info page for any file (by appending &action=info to the URL) will show you what the machine vision platform suggested for that file if it has gone through the Google system (not all files have yet).
There are currently no plans to add the ability to add suggestions, but I'm curious to know how you'd envision possible ways that feature could work.
There have been some high-level discussions about customizing suggestions to the user's interests. On Wikipedias, the WMF Growth team has already started something similar with "topic matching" as part of the newcomer workflows. We could possibly integrate that work into computer-aided tagging when it's ready.
I would stick to the 0 or 1 existing depicts and add some restriction to filter out cases where a very specific item has been added yet. So if it has building (Q41176), suggest more, if it has Grote Kerk (Q1545193), skip it. Probably easiest to look for subclass of (P279)
As a general reminder (also recently voiced by User:Jean-Frédéric): Our only limited resource is user time (community time). Any product should be optimized to use the time spend by the user (and the community as a whole) as effective as possible. You should keep that in mind when working on this tool. For me it has always been one of the leading principles when developing things. Bots do the stupid repetitive and tedious work, community members can focus on the harder things. Multichill (talk) 18:16, 10 September 2020 (UTC)
I also tested the today's version of the tool, opting for my uploads, which usually means my photographs or photographs I imported from flickr or other places, files I cropped etc. Some observations:
A lot of files have very specific categories, like File:Cylindropuntia bigelovii, Cholla Cactus Garden 4837 RobbHannawacker.jpg has "Cylindropuntia bigelovii" in the name and is in category Category:Cylindropuntia bigelovii. However the most specific suggestion is Cylindropuntia fulgida (Q149599). I wonder if one could implement some logic where the software can take ques from filename, description or a category to suggest a tag. At 100% of images of plants, camera models, vehicles, etc. I had to cup and paste some info from the page (from name or category) to add as tag, but suggestions were too broad.
Can I skip "Confirm Tags" step. It is just an extra step that seems pointless. It is like : Publish, are you sure? yes sequence.
When I skip an image because I can not think of a good tag, I have no clue what am I looking at or some other reason, I do not want to see it on that day. I seem to be skipping the same images over and over. I would like to go to the next set of "personal uploads to review".
In case of the artworks with digital representation of (P6243) statement the depict statement about that the artwork depicts should be stored on Wikidata, while on commons we should store info about what the photograph depicts which is usually the same as P6243 and is often omitted.
We are adding a lot of Wikidata items. We need a mechanism for updating statements on commons each time wikidata item id changes. See phabricator:T237899.
This tool is, to me, one of the worst failures in the history of Wiki Commons. I don't know on what basis are the suggestions provided, but they are misleading 99% of time ! I'm importing mostly zoological pictures, and the tool always suggests absurd things (wrong species names [e.g. fish names for mollusks !], wrong vernacular names, wrong biological groups...) whereas all the accurate information is perfectly entered in the description, the categories and the structured data. How can it be ? I fear that good-willed people with insufficient academic background will use this tool to validate inaccurate keywords and create wrong information towards Wiki projects and beyond.
We are already losing a lot of time in entering proper data in all the fields, so please don't add a layer of useless inaccuracy to our work.
There is no discussion on this topic and the tool is still annoying me with senseless tags. I am glad that there is a link on the top of the page how to switch off the tool. It's easy. --Hfst (talk) 16:30, 2 October 2020 (UTC)
Test runs on images uploaded by GLAM institutions
Hi, we would like to do test runs on images provided by GLAM institutions. Would it be possible to create an additional functionality to allow users to tag images from specific Commons categories? - One category we would for example be interested in is: Category:ETH-BIB_Leo_Wehrli. If it turns out to work well, I could well imagine that the semi-automatic tagging functionality would be attractive to GLAM institutions. --Beat Estermann (talk) 15:43, 8 October 2020 (UTC)
Hello @Beat Estermann: . Unfortunately, this is currently not possible with the existing system. We have it in mind for a future update but we can't guarantee when that will happen. In the meantime, for your use case you might be able to repurpose the ISA tool to create sets of images to suggest for tagging. It does not have the machine vision features that Computer-aided tagging has, but it does allow users to enter depicts values manually which may suit your needs well enough. RIsler (WMF) (talk) 21:59, 15 October 2020 (UTC)
Hi @RIsler (WMF): we are interested in the machine-learning part. What can be done to accelerate the implementation of that extra feature that would allow selecting subsets of images for specific tagging campaigns? Are there any particular hindrances that keep you guys from pursuing the project? -- Cheers, Beat Estermann (talk) 09:00, 16 October 2020 (UTC)
@Beat Estermann: we're currently very busy with building the new MediaSearch on Commons. We're a small team and can only handle a limited number of tasks at once, especially as COVID-19, wildfires in the US, and other events reduce our capacity even further. We certainly won't be able to get to this particular feature in 2020. However, the machine vision data on files is available via API calls and if you can find a volunteer developer to put something together we'd do our best to support that effort. RIsler (WMF) (talk) 18:39, 16 October 2020 (UTC)
in 'detailed mode', custom identifications still lack descriptions when added
For example, I looked up 'sleeve' to add to an image that showed only the sleeve of a shirt. I was able to distinguish it from other items with label 'sleeve' because it had a WD description, but as soon as I added it, it was not displayed. Arlo James Barnes22:58, 28 October 2020 (UTC)
Is there something similar for Audio Files
Thats a great feature, and easy to understand. As first time user I checked it ad hoc. Thats cool. Coud you imagine something similar for audio files ? Or ist there something available ? Thanks to the team ! --DrTrumpet (talk) 16:16, 19 November 2020 (UTC)
.jpg)
.jpg){kind=link}
.jpg){kind=link}
.jpg&action=history){kind=link}
_immature_male.jpg){kind=link}
