Artificial Intelligence is coming to PDF accessibility

12th March 2019 | by Ted Page

Most of the things you might do in Microsoft Word to ensure that your documents end up as accessible PDFs will also make life easier for you as the document author, as well as improve the quality of your documents for all of your readers.

The exception—image descriptions 

The most important exception to this rule is the requirement to provide image descriptions, a process that can add considerable time and effort. However, in the latest version of Word, Artificial Intelligence (AI) offers the beginnings of a solution, in the form of auto-generated image descriptions. In some cases (but not all, as we will see) the results can be pretty good.

AI-generated image descriptions

To add alt text to an image in Microsoft Word (2019), right-click the image in question and select Edit Alt Text. The panel below will open. Observe in particular the “Generate a description for me” button.
Screenshot of the MS Word Alt Text dialogue box

A result

In the case of the following image, the alt text generated by Word will be: “A close up of a cat with green eyes looking at the camera”.

A close up of a cat with green eyes looking at the camera
Picture credit Jaroslaw Pocztarski, some rights reserved

The importance of context

It is important to bear in mind that what is appropriate alt text is very much determined by the context in which it appears.

Word’s auto-generated description function appears to look only at the image and not the context. Nevertheless, the result, at least in this case, is pretty good. The cat is looking slightly to one side of the camera rather than straight at it, but otherwise, with a little modification, this description could very well be appropriate in a range of different contexts.

Other examples

However, the results are not always so accurate or useful.

For example, John Constable’s famous painting, Salisbury Cathedral from the Bishop’s Grounds, is described by Word as: “A group of people in a forest”.

Painting by John Constable: Salisbury Cathedral from the Bishop's Grounds

This is clearly not an appropriate description in any context. There is no mention of the cathedral—one of the most iconic buildings in the UK, nor is the landscape by any stretch of the imagination a forest. Nor, of course, is there any mention of John Constable, or that this is a painting and not some other form of image.

Bar charts

Having pasted the following Excel chart directly into Word there was no auto-generated alt text option available (the “Generate a description for me” button was not displayed).

HDI rankings for Afghanistan, Albania, Algeria, Andorra and Angola, 1980 to 2010

However, the same chart converted to a PNG and then inserted into Word returned a description of: “A screenshot of a cell phone”. Note, even if the PNG version had returned a usable description this method is not recommended, as it rasterizes the chart’s axes labels and data values, and thus renders them unreadable by assistive technologies such as Read&Write Gold.

Conclusion

Clearly there is some way to go before this technology will relieve authors entirely of the need to compose accessible image descriptions for their documents. However, as the first example above demonstrates, AI already has the ability to take some of the hard work out of the process of making PDFs accessible, and will undoubtedly continue to do an increasing share of the work in the future.

Watch this space.