Automated accessibility testing
I think it’s fair to say that every application developer needs to take care of accessibility at some point. Indeed, even if you use an accessible toolkit to create an application, the app isn’t always accessible simply because the combination of accessible parts is not always accessible. In reality, many things can go wrong. Without care the app may become completely inaccessible as complexity increases
If you’re a web developer, then you’re (hopefully) already familiar with WCAG, ARIA and other cool stuff helping to address accessibility issues in web apps. If you are a platform developer such as an Android or iOS, then you probably know a bazillion of tricks to keep mobile apps accessible. You might also know how to run the app over screen readers and other assistive technology softwares to ensure everything goes smoothly and works as expected. And you might already have started thinking of how you can use automated testing to cover all the accessibility caveats to avoid regressions and to reduce overhead of the manual testing.
So let’s talk about accessibility automated testing. There’s no universal solution that would embrace each and every platform and every single case. However there are many existing approaches, some are better than others. Some of them are fairly solid. Having said that looking at the diversity of the existing solutions and realizing how much people still do the manual testing (AAM specs, such as ARIA or HTML accessibility APIs mappings, can be a great example of it), I think it’d be amazing to systemize the existing techniques and come with a better, universal solution that would cover the majority of cases. I hope this post can help to find the right way to do this.
What to test
First things come first: what is the scope of automated accessibility testing, i.e. what exactly do we want to test?
Without a doubt, the web is vast and complex and plays a significant role in our lives. It surely has to be accessible. But what is an accessible web, exactly?
The main conductors on the web are web browsers. They make up the web platform by providing all of the tiny building blocks such as HTML elements to create web content or ARIA attributes to define semantics. Browsers deliver web content to the users by rendering it on a screen and expose its semantics to the assistive technologies such as screen readers. All these blocks must be accessible. In particular that means the browsers are responsible to expose all building blocks to the assistive technologies correctly. It’s very tempting to say if a browser is doing the job well, then a web author cannot go wrong by using these blocks (for sure, if the web author is not doing anything strange on purpose). It sounds about right and it works nicely in theory. But, in practice accessibility issues suddenly pop up as the complexity of a web app goes up.
The browsers are certainly the major use case in the web accessibility testing and I’d like to get them covered, simply because the web cannot be accessible without accessible browsers. Also, they already do a decent job on accessibility testing, and we can learn a lot from them. Indeed, they’re all stuffed with accessibility testing solutions and each has its own test harness for automating the process. Their system could be unified and adjusted to a broader range of uses.
The web applications are the second piece of the web puzzle. They also must be covered.
The webapps are made up of small and accessible building blocks (if the browser does a good job). However, as I previously noted, it is not sufficient to have individually accessible parts — it is necessary that all of the combinations of parts are accessible. It may sound quite obvious since QA and end to end testing wasn’t invented yesterday, but this is something overlooked quite often. So this is one more yet use case on the table: the overall integrated web application’s accessibility.
Although the web is vital, there’s a sizable market of desktop and mobile applications which also need to get accessibility tested. Some desktop/mobile apps are using embedded browsers under-the-hood as a rendering engine, which brings them to the web scope. But speaking generally, desktop/mobile apps are not the web.
Having said that, it’s worth noting that browsers and desktop/mobile applications coexist in the same environment, and they use the very same platform accessibility APIs to expose the content to the assistive technologies. It makes web and desktop/mobile applications have more in common than people usually tend to think. I’m from the browser world and I keep focus on the web accessibility as many of you probably are, but let’s keep in mind desktop and mobile applications as one more use case. We will have them covered as well.
How to test
The next question to address is how to test accessibility or how to ensure that the application is accessible. You could test your entire application with a screen reader or a magnifier on your own and that would be pretty trustworthy, but what exactly to test when it comes to automated testing?
Unit testing is the first level of automated testing, and accessibility is no exception. It allows you to perform very low-level testing such as testing individual c++ classes or JS modules, or testing all internal things that are never disclosed or can never be tested via any public API as the pure under-the-hood thing. As a result, unit testing is a critical component in automated testing.
However, it has nothing to do with accessibility specifically, and there’s nothing to generalize or systemize for the benefits of accessibility. It’s simply something that all systems must have.
Accessibility APIs mappings
When it comes to automated accessibility testing, the first and probably the best thing to start from is to test accessibility APIs. Accessibility APIs is the universal language that any accessible application can be expressed in. If accessibility properties of UI looks about right, then you have great chances the app is accessible. By the way, this is the most common strategy in accessibility testing today in web browsers.
To give an example how accessibility APIs testing can be used practically, you can think of AAM web specifications. These specs define how ARIA attributes or HTML/SVG elements should be mapped to the accessibility APIs. This is a somewhat restrictive example to demonstrate the capabilities of accessibility APIs testing, because AAM utilize only a few things peculiar to accessibility APIs (for example, it misses key things like accessible relations, hierarchies or actions) but it’s a good example to get a sense of what kind of things can be tested when it comes to accessibility APIs.
Accessibility APIs have complex nature. They differ from platform to platform, but they share many traits in common. It’s not surprising at all, because they all are designed to expose the same thing: the semantics of user interface.
Here are the main categories you can find in many accessibility APIs.
- States and properties: an example for these could be focused or selected states on a menu item or accessible name/description for properties.
- Methods or interfaces allow to retrieve complex information about accessible elements, for example, to retrieve information about list or grid controls.
- The relation concept is a key one. It defines how accessible elements relate to each other. In particular they are used to navigate content and/or to get extra information about elements such as labels for a control or headers for a table cell.
- Accessible trees are a special case of accessible relations. However they are often defined as a separate entity. The accessible tree is a hierarchical representation of the content, and the ATs frequently rely on it.
- As a rule of thumb there is also special support for text and selection.
- There are also accessible actions for example a button click or expand/collapse a dropdown
- Accessible events are used to notify the assistive technologies about app changes.
All of this is a very typical if not comprehensive list of what can be (or should be) tested when it comes to accessibility platform APIs testing.
Accessibility APIs are fairly low level testing, which is good or might be not depending on a use case. I would say it’s nearly perfect to test browsers, for example, how the browsers expose HTML elements and such. It also should be the right choice for different kinds of toolkits, which provides you with a set of building blocks, for example, extra controls on top of HTML5 elements. I’m not quite confident that this type of testing is exactly what web developers look for when it comes to webapps accessibility testing, because it’s fairly low level and requires understanding of under-the-hood things, but certainly they can benefit from checking accessibility properties on certain elements/structures, for example, for web components.
Assistive Technologies testing
Assistive Technologies (AT) testing makes another layer of automated testing. Unlinke accessibility APIs testing, this is a great way to ensure your app is spoken exactly as you want over different screen readers or it gets zoomed properly when it comes to screen magnifiers. Any application including both browsers and web apps can benefit from integrational AT testing by running individual control elements or separate UI blocks though the ATs.
AT testing can be also used for end-to-end testing. However, as any end-to-end testing it cannot be comprehensive because of a steep rise in testing scenarios as the app complexity increases. Having said that, it certainly can be helpful to check the most crucial scenarios and can make a real great addition to the integrational accessibility APIs testing.
Testing a single HTML page to check an accessible tree and/or its relevant accessibility properties makes a nice pattern of atomic testing. It is quite similar to unit testing. However, the reality is that not every use case can be reduced to a simple static page. It makes such testing quite restrictive. The real world affirms we need to test accessibility in dynamics for a full spectrum of testing scenarios, such as clicking a button and checking what happens next.
This makes another important piece of the puzzle, a testing flow, or in other words, how to control an application to test its dynamics, where you can query accessibility properties at the right time.
A typical test flow can be described in three simple steps:
- Query-n-check: test accessibility properties against expectations.
- Action: change a value or trigger an action: it represents the dynamics part and describes how a test interacts with an application.
- On-hold: wait for an event to hold the test execution until the app gets to the right state.
To summarize there are two key pieces we’d need for a testing flow. First, to trigger an action. This can be emulation of user actions, triggering accessible actions or changing accessible properties, anything that allows the user to operate and control the application. Secondly, we need the ability to hold on to a test execution until certain criterias are met. That could be an accessible event or presence of certain accessible properties on an element, or waiting until a certain text is visible, whatever, just to put the test execution on pause until it’s good to proceed.
What we have now
Let’s take a look at what the market has to propose. I don’t know much about desktop or mobile applications, many of them are not open source, so no chance to sneak a peek. But I think it’d be a good idea to start from the web and the web browsers in particular, who have made some real good progress on accessibility testing.
The first thing I would like to mention is AOM. It’s not the browsers, but it’s something closely related. AOM stands for Accessible Object Model. This is an attempt to expose accessibility properties in a cross-platform way in browsers. Similarly to platform accessibility APIs, AOM defines roles, properties, relations, and an accessible tree. You can think of AOM as the web accessibility API. So having AOM with all typical accessibility features can make a great platform for cross-browser accessibility testing.
Certain elements like platform dependent mappings are not very suitable for AOM for sure. For example, AAM specs, the accessibility APIs mapping, which define how to expose the web to the platform accessibility APIs, do not make a great choice for AOM testing. Indeed, despite platform APIs similarities on the conceptual level, they differ in details. However, if we care about the logical compound only, AOM will suit nicely. For example, ARIA name computation algorithms make a great match for AOM testing, or AOM can be used to test relations between a label and a control. In this case we don’t worry what is the platform name for the relation, the only thing we want to test is the right relations are exposed.
Sadly, it’s not something that was implemented. AOM is a long term vision, and the main focus till now has been on ARIA reflection, which is prototyped by a number of browsers by now. But ARIA reflection spec is just a DOM reflection of ARIA attributes and thus has somewhat lower testing capabilities. For example, HTML elements that don’t have a proper ARIA mapping cannot be tested by ARIA reflection or any of accessibility events.
So AOM is something that has great testing potential but it’s not yet implemented or even specified.
ATTA (Accessible Technology Test Adapter API) was the answer to manual accessibility testing for ARIA and HTML AAM specifications. Roughly speaking ATTA defines a protocol used to describe accessibility properties for a given HTML code snippet to create the test expectations. The expectations are sent to the ATTA server, which queries an ATTA driver for an accessible tree for the given HTML snippet and then checks it against the given expectations.
The ATTA is integrated into WPT test suite and it could make a great solution for the web accessibility APIs testing, if it worked. There are implementations of ATTA drivers for IAccessible2, UI Automation and ATK but apparently none of them ever reached the ready-to-use status.
So ATTA has a bunch of worthy ideas like WPT built-in integration and the modular system which allows you connect various platform APIs drivers, but sadly it is not finished yet and no active work any longer.
Let’s take a look at the browsers. Browsers have quite a long history of accessibility support and they’ve made a great progress on accessibility testing.
Here’s an example of the Gecko accessibility test. The test represents a typical scenario for the accessibility testing: you get a property, call an action, wait for an event, and then make sure the property value was adjusted properly. You can imagine that cross platform tests are quite similar to this one.
Gecko implements its own test suite which provides a number of util functions such as ok() or is(), which are responsible to handle and report success or failures. This is the kind of testing system, where test expectations are listed explicitly in a test body, or in other words, the test says what to test and what are expectations. As a direct consequence if you need to change expectations, you have to adjust the test manually. It’s quite a typical testing system though.
The testsuite itself is quite different from Firefox though, which is not surprising. The WebKit test generates output which is compared to expectations stored in files. The test expectations are also listed in a test body, which makes the approach close to Gecko.
Similar to Gecko, WebKit also supports event testing, here’s an example of a typical test. It might look bulky for a simple thing it does, but all stuff can be wrapped by a nice promise-based wrapper which will make the test more readable. The most important thing here is that WebKit also supports testing for all parts of a typical accessibility API.
Beyond low-level c++ unit testing, Chromium relies on platform accessibility APIs testing. It is quite similar to Firefox or WebKit with one key difference.
Chromium can dump an accessibility tree with relevant accessibility properties or can record accessibility events, which it subsequently compares to the expectation files. The main gotcha here is these tests can be rebaselined easily unlike other kinds of tests. If something changes on the API level, for example, if an accessible tree is spec’d out differently, new properties are added or old ones are removed — any change, then all you need is to rerun a test and capture output, which becomes the new expectations. It’s as if you take a snapshot that becomes a new standard, and then all following runs match it.
Here’s a typical example of a tree test with mac and win expectation files.
Chromium also reveals basic scripting capabilities. Those are mainly mac targeted though, and thus scoped by NSAccessibility protocol testing. However it allows testing all bits of the API, including methods, attributes, parameterized attributes, actions and events.
What would make a great test suite?
Let’s pull things together. Accessibility exists within the context of a platform, which glues together an application and the assistive technology by accessibility API. We want a platform-wide testing solution to test platform accessibility APIs. It should be possible to test a variety of things such as accessible trees, properties, methods and events.
The solution should not be limited to just web browsers. It should be capable of covering web applications running in web browsers. We’d also like to test any application running on the system.
Multiple platforms should be supported such as AT-SPI/ATK on Linux, MSAA/IA2 and UIA on Windows and NSAccessibility protocol on Mac. The solution has to be extensible in case we need to support other platforms in future.
Test flow control should be supported out-of-the-box, such as interacting with an app and then waiting for an accessible event. Another example of interactions will be the flow control directives allowing communication with the assistive technologies. Getting those covered will allow writing end-to-end tests.
Last but not least. Easy test rebaselining is a key feature to have. If you ever need to change a test’s expectations, then you just rerun the test and record its output. It happens more often than you probably think. This testsuite allows you to adjust expectations with about zero effort.
Chromium accessibility tools
Chromium has decent accessibility tools to inspect a platform accessibility tree and to listen to accessibility events. They are available on all major platforms and can be easily ported to other platforms, essentially on any platform Chromium is running on. They are capable of inspecting any application on a system including web applications running in a web browser. All major desktop APIs are supported as well, namely AT-SPI on Linux, MSAA/IA2 and UIA on Windows and NSAccessibility protocol on Mac.
In Chromium these tools are integrated into a test harness to perform platform accessibility testing. The testsuite supports test flow instructions and rebaselining mechanism.
The tools can be beefed up with testing capabilities to become a new test harness, or be easily integrated into existing test suites or CDCI systems.
The tools are not perfect and are not feature-complete. They have great potential though. As long as they are capable of providing all the must-have testing features we discussed above, all they need is some love I think. The tools are open source and anyone can contribute. However, having them as an integral part of the Chromium project makes it seem like they are inherently tied to that project and doesn’t make life easy for new contributors. It’s possible, however, that eventually the tools could get a new home. If new contributors bring the fresh vision and new use cases to shape the tool’s features, then it can make a great start of a new open source project which will embrace all innovations, and hopefully can solve a long standing problem of accessibility automated testing.
Is it worth taking a shot?
Are you ready to join the efforts?