ML Applications for Web Accessibility: Automated Testing for Interactive Content Elements, Images and Text
This article covers insights on web accessibility in an era of User Experience
Web accessibility is the practice of making digital products, such as websites, more accessible and navigable for all users, including those with disabilities and impairments. Much like designated parking and service ramps, European and U.S. governments have included the internet as a “place of public accommodation,” which must meet accessibility guidelines. Web accessibility covers a broad range of limitations – whether someone is blind or has a broken arm – and accessibility features seek to ensure the same user experience regardless of disability, injury, or impairment.
Top AIT Interview: AIT Interview with Aaron Kalb, Chief Data & Analytics Officer at Alation
The genuine necessity of web accessibility cannot be understated because it allows users with diverse disabilities to effectively navigate, understand and use platforms, content, and tools without compromising user experience – which is everything in our modern world. Proper disability accommodation encompasses all digital products, including websites, mobile device applications and educational resources.
Nevertheless, making any digital product accessible is a meticulous process, and that is why experts are looking towards machine learning (ML) as a potential solution.
WCAG Standards and Automated Accessibility Testing
Web accessibility relies upon accessibility testing, which checks that assistive devices can translate and convey web material into information that a person with an impairment could easily understand. Essentially, the testing checks digital products for compliance with standards laid out by the Web Content Accessibility Guidelines (WCAG), or Section 508 in the USA, to be exact. Digital products must follow version 2.0 of WCAG, at least, which defines how to make web content accessible to people with visual, auditory, physical, speech, cognitive, language, learning and neurological disabilities.
When testing web content for accessibility, it is relatively easy to automate requirements for the markup language that defines the structure and presentation of the content. However, this only accounts for about 20% of common accessibility support issues per WCAG 2.0/2.1 level AA. The other 80% primarily involves issues concerning images, audio and video, interactivity and meaning of the text. Usually, these issues would be manually checked, resulting in a repetitive, labor-intensive and time-consuming process. But now it is possible to automate this process with ML.
Different content types require different testing models involving various ML disciplines. Natural Language Processing (NLP) models are typically used for checking the correctness of the page structure and navigation, but also used in almost all other ML automated tests as their integral parts. Computer Vision (CV) models are most commonly applied for checking contrasts, alt texts and finding groups of elements. Signal/audio analysis can help automate testing for checking audio description, audio and video captions and transcripts.
For example, one of the WCAG criteria for accessibility is the requirement to include alternative text for pictures, graphics, and other similar content. This text is used by screen readers to describe digital images verbally for people with limited vision. Often, text alternatives are not included, or such text is incomplete or not informative, which is a serious violation of WCAG.
A combination of CV and NLP techniques can be used to validate alternative text for images. The algorithm automatically interprets the image and then compares the generated annotation with the text and adds it to the picture or graphic as the text alternative; the two texts get evaluated for their degree of similarity. After that, an algorithm will conclude whether there was a WCAG violation or if the image has the correct alternative text.
Another common WCAG requirement (and an equally common accessibility violation) concerns the ease of site navigation for users with limitations. Accessibility standards entail that each web page has a descriptive title, clear headings and labels, easily understandable links, etc. Confirming these requirements is time-consuming when done manually, but with automated NLP models and the semantic processing and comparison of keywords and texts, it is much quicker.
For example, to check whether the description of a link matches the page content to which it leads, keywords are extracted from both texts (the link description and the page content). Selecting the most appropriate keyword involves various statistical approaches: word frequency, collocations, and co-occurrences, and processing other text statistical features. Finally, using text similarity techniques, for instance, a combination of different word embeddings and cosine similarity, the keywords from both documents are compared. Similar approach is used for confirmation that the title, headings and labels match the page content.
The Main Challenges for Automation of Accessibility Testing with ML
Checking of the arbitrary content.
When setting the task, initially, you need to acknowledge the fact that the content is constantly updated. Besides, the layout of website pages is not always unified and there are a lot of ways to embed web content. Moreover, many of the frameworks on which content is produced are un-opinionated. This means that the developer can make independent decisions and has more control over how he codes different functionality. Consequently, this affects how the web content is displayed, and therefore, perceived. This leads to a wide variety of possible implementations of the same web elements. Therefore, accessibility testing automation should initially support validation of any type of web content design and structures. This implies the presence of many examples to train the ML models used. Thus, a large amount of well-prepared and labeled data is required.
Coordination of multiple ML models for a single accessibility test.
Another problem is that accessibility checks require multiple steps from the accessibility tester. For example, WCAG 2.4.2 criteria require that each web page has a descriptive title. The tester must find the title element on the page, read it, read the page content, and match the title and information on the page. He needs to make sure that the title reflects the content of the page and to decide how much it matches. If a machine performs this test, then separate models and algorithms are responsible for each individual step. Thus, performing such checks requires a combination of different algorithms, models, and business logic. The development and maintenance of such combined solutions require the well-coordinated work of several specialists: programmers, data scientists and data engineers all at once.
How ML Makes Accessibility Testing Streamlined and Cost-Efficient
Although accessibility testing is a labor-intensive endeavor, ML can streamline and simplify many challenging processes for cost-efficient results. The automation of testing for interactive content elements, images and text can increase the volume of detected issues according to WCAG 2.0/2.1 levels A and AA criteria up to 70% and decrease human labor by 40-90%. Similarly, ML can support audit reporting and automatically produce various Voluntary Product Accessibility Template (VPAT) reports.
Additionally, when an error occurs in the accessibility process, the costs of fixing that error – which usually get overlooked until the final stages of development – can be expensive. But, by deploying an ML solution, the algorithms can detect issues, commonly accessibility gaps, in the early stages of development, eliminating the costs associated with fixing an error later. Likewise, ML configurations can offer specialized recommendations for fixing problems and confirm that all code changes work as intended, streamlining the responsibilities of quality assurance teams.
As new regulations get rolled out by WCAG, ML emerges as a real solution to the challenges of accessibility testing. The automation possibilities offered by ML are cost-efficient, diminish the need for manual labor and can bring the most value when applying the shift-left approach to accessibility verification in the CI/CD process.
[To share your insights with us, please write to email@example.com]