Wikimedia Apps/iOS おすすめ編集プロジェクト/Alt 文の実験
背景
これに取り組む理由は?
画像はますます普及しており、情報の提供ができる点: 直近の調査によると、ウィキペディアの記事に置いた画像は閲覧者がその記事に関与するレベルを引き上げ、挿し絵の役割に加えて情報も提供します。[1] If a user cannot view the image due to low or no vision, or internet connectivity, they can utilize the alternative text (or alt text). Alt text is text that can be associated with an image that serves the same purpose and conveys the same essential information as the image.[2]
ウィキペディアの画像で代替文がないものが多数あります: 2021年の調査では、英語版ウィキペディアの画像の 46% にキャプションが、同じく 10% に代替文が付いていて、有効な代替文は画像全体のわずか 3% にしか付いていませんでした。[3] Addressing the lack of Alt Text of images in Wikipedia Articles, and on Commons has long been a topic of discussion, research, and organizing. Various alt text editing events have been organized by Affiliates in Poland, Ireland, Argentina, and through global events such as Image Description Week 2022.
ウィキペディアのアクセス性を高めること: Recommendation #2 from the Movement Strategy is to improve user experience by supporting compliance with the most advanced accessibility guidelines using free and open-source software (WCAG for web, W3C mobile web best practices, etc.). A recent study of image accessibility on Wikipedia across languages suggested that one part of the solution is to invest in tools that can surface articles and images without accessibility coverage.[4]
The iOS Wikipedia app was Apple's editor's choice in 2017 as a result of the app's user accessibility features. Accessibility is an important factor to our design and development process on the apps teams, so a task to fill the gap in images with quality alt text is fitting for our team.
Our early prototype showed promise:We created proof-of-concept experimental build for adding alt text to images on Wikipedia in the context of their article, and shared it with users at the 2023 GLAM conference.
After getting feedback, we feel confident with scaling our experiment in the production version of app. Part of this experiment builds on top of our recent addition to the iOS app of the "Add an image" suggested edit.
User stories
- As a Spanish Wikipedia editor in LATAM using the Add an Image feature for the first time on the Wikipedia iOS app, I would like to build confidence to add alt-text while on the bus, so that I can add alt-text to future image related tasks without concern of making the state of alt-text worse on Wikipedia for low vision users.
- As a Portuguese Wikipedia editor in Brazil, I would like to become aware of opportunities to add alt-text to images in articles I care about, so that I can ensure all editors are able to gain the full context of those articles
- As a user navigating articles with a screen reader, I want quality alt-text to be available , so that I have the same additional context about an article as users that are not using screen readers
- As a user with limited data, I want to read alt-text, so even if images are not loaded, I am aware of what is in the image
Objective
If we conduct an A/B/C test with the alt-text suggested edits prototype in the production version of the iOS app we can learn if adding alt-text to images is a task newcomers are successful with and ultimately, decide if it's impactful enough to implement as a suggested edit on the Web and/or in the Apps.
This work is part of the 2024-2025 annual plan Wiki Experiences 1, focused on Contributor Experience.
The scaled experiment will not be a feed of images in need of alt text. Rather, we will prompt users to add alt text to images after they have made a related edit: either adding the image using image recommendations, or after editing an article. If the user accepts the prompt, they will be taken through a dedicated flow to add alt text to the image, with guidance and examples.
Through this experiment, we have the opportunity to
- Learn if newcomers and experienced editors can successfully add alt text through a structured task with guidance
- Try out 2 different moments to suggest additional edits to users:
- After a user completes a suggested edit
- After a user completes an article edit
- Increase the number of images on target Wikipedias that contain useful alt text
Experiment Design
Research questions
- Does reminding users to add alt text prove an effective way to increase the number of images with alt text, and number of "Add an image" edits that include alt text?
- Should adding Alt text for images in articles be a stand-alone suggested edit? (And if so, would it be an appropriate task for newcomers?)
- Would users appreciate prompts for other improvements they can make to the article they are editing?
- Does a guidance help newcomers and experienced editors complete the task?
Experiment B: Add alt text Prompt after "Add an Image" flow
Currently, only around 16% of edits completed through Image Recommendations add alt text. We want to learn if reminding users to add alt text prove an effective way to increase the number of edits that include alt text. Editors will only be shown the prompt once.
- Hypothesis: if we prompt users once to add missing alt text to their image after they have published an image using the Image Recommendations suggested edit, we'll see 60% of editors choose to additionally publish alt text, and 10% will add alt-text for subsequent image recommendation edits made in the next 15 days.
- Audience: logged-in editors have completed an edit in the "Add an image" feature where alt text = null, and who have not already been sorted into experiment group C
- Group A = 50% of "Add an image" editors are sorted into Control group when they forget alt text. They are not prompted to add missing alt text.
- Group B = 50% of "Add an image" editors are sorted into Experiment group when they forget alt text. They are prompted once once to add missing alt text
- Assignment into group A/B happens when a user completes an "Add an image" suggested edit where alt text = null.
Experiment C: Add alt text Prompt after standard editing flow
We want to learn if users would appreciate prompts for other improvements they can make to the article they are editing. This is a type of in-time suggested edit. Editors will only be shown the prompt once.
- Hypothesis: if we prompt users to add missing alt text after they have published an edit on any article containing an image in need of alt text, 4% of editors will go on to add alt text to the image.
- Audience: logged-in editors who made any type of edit on an article with an image lacking alt text, and who have not already been sorted into experiment group B
- Group D = 50% of editors are sorted into Control group, they do not see any prompt after completing an eligible edit
- Group C = 50% of editors are sorted into Experiment group, they see 1 alt-text prompt after completing an eligible edit.
- Sampling into group D/C should happen immediately after a user publishes an edit that’s eligible for the prompt.
- Eligible edits for follow-up prompt: any edit made by a logged-in user on an article in the main namespace
Feature Requirements
Must haves
- Entry point will vary based on experiment group (see below)
- If a user answers "no" to "Add alt text" prompt, gain reason for why
- Prominent guidance for writing good alt-text
- Users ability to get context about the image from the article
- Users ability to access relevant metadata (can take user to Web)
- Detection of which images do not have alt-text
- Ability to publish alt-text
- Alt-text edit should count as a separate edit ONLY when submitted separately
- Alt text is published with automatic edit summary
- Users should be prompted provide feedback about the feature
- Instrumentation that allows us to evaluate the alt-text submitted
- Input field for alt-text in context of image, preview of alt-text in context of our existing editor
- Warning when alt-text exceeds 125 characters
- Entry field should not allow users to add line breaks
Nice to have
- Positive Reinforcement
- Do not allow users to copy and paste in image caption
- Suggest Alt-text (think Machine Assisted Article Descriptions)
- Can playback what was written in preview
- Surface Categories and Depicts
Experiment B Must Haves Experiment Group B (50% of Image Recs Suggested Edits users). They will enter the dedicated flow for Alt Text AFTER adding an image to an article using Image Recommendations.
Entry point: After image recommendations edit is submitted; if alt-text = null:
- Educate user on importance of alt-text and ask if they’d be willing to add alt-text to image
- If user selects yes; launch alt-text adding flow for most recently added image
After task completion:
- Show a survey when to ask users if it should be a separate task
- Return users to next image recommendations suggestion
Experiment C Must Haves Experiment Group C: 50% of all editors receive encouragement to add alt-text to an image when editing an article with an image
Entry point:
- After edit where either of the following were true
- No edit was made to an image, but there is an image in the article without alt text
- Edit made to image but alt-text = null
- Show prompt that an image in the article is in need of alt-text and explain importance of adding alt-text and ask users if they’re willing to add alt-text
- If users selects Yes, launch alt-text adding flow for one image in article in need of alt text
After task completion
- Ask satisfaction
- Ask if this should be a dedicated task
- Bring user back to article they were reading
Experiment C Nice to Haves
- Affordance to reveal / review alt-text on images in an article
- If someone is abandoning an edit, launch the prompt for adding an image if they have over 50 edits
Target Wikis
While we welcome feedback from everyone, we are especially interested in hearing from:
- Spanish, French and Portuguese speakers in the Americas (North, South and Central) and Caribbean
- Chinese speakers in North America
We are planning to run this experiment in partnership with the following wikis:
- スペイン語版ウィキペディア
- フランス語版ウィキペディア
- ポルトガル語版ウィキペディア
- 中国語版ウィキペディア
How will we know we are successful?
Evaluating the quality of Alt Text:
In our preliminary research, we learned that alt text additions are not formally patrolled, so we may not be able to rely on revert rate alone to understand the quality of alt text being added. We plan to partner with an accessibility organization to evaluate the quality of the alt text produced through this experiment. We will ask them to rate each alt text entry from 1-5 with 5 being the highest quality, and 1 being lowest quality. They will also provide a predicted revert score, answering the question: if they saw that alt text added to an image, would they remove it from Wikipedia?
Leading indicators to be measured after 15 days
- 100 edits with alt text values, from at least 25 unique editors. At least 25 edits are from newer editors
- More than 15 unique editors have been assigned to each experiment group
- 70% task acceptance rate for group B, at least 10% acceptance rate for group C (# of people who enter the flow / impressions of prompt)
- Revert rate for newer editors edits in any single group does not exceed 18%
Key Indicators
- 60% of group B editors publish an additional edit with alt text for the image they were prompted on
- 4% of group C users add alt text when prompted after editing an article
- Of group B editors who make a subsequent image recommendations edit in the next 15 days, 25% add alt-text as a part of that edit
- 200 images are improved with Alt text, by at least 50 unique editors
Decision Matrix for Next Steps
- If 71% of edits are scored a 3 or higher* we will scale the feature. If less than 70% of edits are scored a 3 or higher we will improve guidance or use AI to better assist users.
- If quality scores* for newer editors are more than 50% worse than quality scores for experienced editors, we will not recommend this task be available to newer editors.
- If we see at least 60% say they would use feature that provided a feed of images in need of alt text, then we will have the confidence to pursue a feed of alt-text suggested edits
- If 60% or more of respondents say they would be interested in similar edit notifications for articles they are working on, and 60% of respondents are satisfied with the feature (Group C survey responses), we share this information and consider future edit prompts.
Guardrails
- Edit return rate of editors in group B or C who have received an Alt text prompt does not differ from controls by more than 10%
- Revert rate for edits from experiment groups does not exceed controls by more than 5 percentage points
- Human-graded* or actual Revert rate for newer editors in experiment groups does not exceed 18%
- Alt text task completion rate for newer editors is above 25% (Completion rate = number of alt text edits published / those who said “yes” to the prompt and started the flow)
Curiosities
- How does the task completion rate and revert rate for newer editors' alt text edits compare with that of experienced editors? With that of comparable rates from Growth suggested edits?
- How does the human-graded* revert rate compare to Android's Image Captions Suggested Edit?
- Is there a difference in metrics by language and geography? (For example, breaking down edits from Latin America vs Europe for Spanish)
- What is the most common reason that users decide not to act on the prompt?
* Note: for the quality scores and human-graded revert scores, we will partner with an accessibility organization who will be reviewing and grading alt text.
Definition of newer editors: Editors who had fewer than 10 edits on that wiki they are currently editing at the point they entered the experiment
Designs
Experiment B (Image Recommendations Flow)
-
Add image?
-
Add image details
-
Add caption
-
Preview
-
Edit summary
-
Edit confirmation
-
Prompt to add alt text appears if user has forgotten to add alt text to previous image
-
Asked for reason if they close or choose "Do not add"
-
Onboarding guidance on how to write alt text
-
Tooltip: Review
-
Tooltip: Inspect
-
Tooltip: Add description
-
Add alt text from a halfsheet, with the article behind
-
Halfsheet can be minimized to review article
-
View image information
-
View image fullscreen
-
Typing alt text
-
Warning for exceeding 125 characters
-
Review entered alt text and caption
-
Feedback on suggested edit
-
Feedback submitted and returned to next image recommendation
Experiment C (Article Editing Flow)
-
Publish confirmation from article edit
-
Prompt to add alt text to an article image
-
Onboarding
-
Add alt text sheet
-
Flow continues as shown above in Experiment B (typing alt text, review, publish)
-
Survey on Satisfaction
-
Survey on receiving similar notifications about suggested edits
-
Feedback submitted and returned to article view
How to follow along
We have created T357437, Alt-Text Suggested Edit Scaled Experiment on iOS as our Phabricator epic. We encourage your collaboration there or on our Talk Page. There will also be periodic updates to this page as we make progress on the experiment.
更新情報
2024年8月 Special Update
- The alt text experiment has been released to Beta Testers, on Spanish, Portuguese, French, and Chinese Wikipedias. Editors making edits on articles containing images without alt text, or adding images using “Add an Image” may see one prompt asking them to add alternative text to an image. The experiment will run until early November.
- Check out a demo of the flow on Portuguese Wikipedia:
Two A/B tests running in the iOS App that prompt users to add alternative text after a relevant edit - If you would like to download the beta version of the app to test, please follow these testing instructions.
2024年7月
- We held a Deep dive meeting with the team and stakeholders from the GLAM and Research teams to review the experiment plan, designs, and timeline. Most recent designs have been added above
- We defined our comparison group for learning how newer editors performed on the task as: editors who had fewer than 10 edits on that wiki they are currently editing at the point they entered the experiment. We’ll use the term “newer editors” to refer to this group, as not to confuse them with brand new accounts.
- The team began development on the Alt text experiment, you can follow our work on the Engineering Epic: T357440
- We learned that in iOS, the baseline for image recommendation edits that contain alt text is 16%.
- We reviewed a sample of the alt text that had been entered by users in Image Recommendations. We saw that 15% of edits evaluated repeated the exact same text in the image caption as in the alt text, and that only 14% of edits evaluated had alt text of reasonable quality. This stresses the importance of adding guidance for writing effective alt text in our flow, and for preventing users from copy / pasting between the fields.
- We learned about a new form of AI that is creating alt text for images on the London Museum’s website. If results from this experiment show that users struggle with writing alt text, this could be an avenue to explore for adding AI assistance.
2024年6月
- We completed analysis of usability testing for flows B and C in English, and will be updating the designs based on these results and design review. T360567
- Summary of results:
- There was clear understanding on what the alt text prompt was asking them to do
- For flow B: 5/5 understood that they were being asked to add alt text to the previous image because it's missing
- For flow C: 5/5 understood that they were adding alt text for an image from the article you just corrected the misspelled word in.
- The onboarding, tooltips, and guidance were clear and helpful to users, with some room for small improvements
- 10/10 thought the onboarding was clear
- 9/10 thought the tooltips were clear and helpful
- Users would be able to zoom, but their first choice was not supported
- 5/10 tapped on the image first, then pinch and zoom
- 5/10 would pinch and zoom first on thumbnail, then tap to open
- Most users could find the guidance they needed
- 6/10 would open the "Guidance for writing alt text" link
- Users could find additional information about the image
- 6/10 knew to click on the info-i next to the image, 1/10 found it after being prompted.
- 1/10 tapped on the image first, then found the info-i button -Most testers thought the image information provided was complete, while others noted room for small improvements
- 6/10 felt the information was complete and not missing anything important
- Testers understood the task, and were able to successfully summarize it afterwards
- All testers felt confident they could add alt text to a new image
- There was clear understanding on what the alt text prompt was asking them to do
- Summary of results:
2024年5月
- 利用者テストの分析は、英語版と中国語版で進行中です。内容はAlt文を追加する手順のプロトタイプ2種を用いました。
- 手順Bでは「画像を追加」拡張機能を使った利用者に、その画像にAlt文を追加するよう促しました。
手順BをAlt文プロトタイプの実験に用いた - 手順Cでは利用者が誤字修正などの編集をした記事を対象に、Alt文を追加するよう促しました。
手順CをAlt文試験プロトタイプの実験に用いた
- 手順Bでは「画像を追加」拡張機能を使った利用者に、その画像にAlt文を追加するよう促しました。
2023年9月 – 2024年4月
For earlier updates about the Alt Text suggested edit, see the overview page's updates.
Below are designs for the experiment as of April 2024:
-
Add missing alt text modal
-
Onboarding tips (5 points)
-
First tooltip: Review article
-
Second tooltip: Inspect image
-
Third tooltip: Add description
-
Initial view to add alt text
-
Info-i icon leads to image information from common's listing
-
Inputting alt text
-
Preview screen with alt text on image
-
Automatic edit summary is added, confirmation after publishing
References
- ↑ https://epjdatascience.springeropen.com/articles/10.1140/epjds/s13688-021-00312-8
- ↑ https://www.w3.org/TR/2008/NOTE-WCAG20-TECHS-20081211/G94
- ↑ https://techblog.wikimedia.org/2021/09/09/the-wikipedia-image-caption-matching-challenge-and-a-huge-release-of-image-data-for-research/
- ↑ https://wikiworkshop.org/2023/papers/WikiWorkshop2023_paper_25.pdf