Making science, technology, engineering, and math (STEM) education available to all who are capable of learning is a moral imperative. Yet our educational system (including authors, textbook publishers, and college professors) struggles to provide appropriate access to the images and figures that are critical to STEM subject learning.

Morash et al. envision a system that would allow nonexpert workers (recruited through a service such as Amazon’s Mechanical Turk) to create high-quality accessible descriptions (also known as alt-text) of STEM images. The success of such a system could greatly reduce the cost of making STEM teaching materials accessible, and thereby greatly increase access to STEM education for people with visual impairments.

Their main contribution is to demonstrate that the design of the system’s user interface influences the completeness and uniformity of the resulting alt-text. Current web-based systems for this problem simply present the worker with an image and a set of instructions, and allow the worker to enter his or her description as free text. The authors have created a competing system (which they call a queried image description, QID) that uses an interactive survey tool to gather information from the worker, and then auto-generates the image description using a template. Web workers using QID are significantly less likely to omit key information (like captions, or units on graphs) than those using free text entry. There is also significantly less variation in descriptions generated by different workers when using QID as compared to free text entry, which should simplify quality control and reduce user confusion.

While there is a well-founded hope that QID-generated alt-text will be more usable than free text, and comparable to alt-text created by experts, such usability testing is left as future work.

Another notable aspect of this paper is the bringing together of “greatest hits” from several different areas of computer science research. Ukkoken’s approximate string matching, Von Ahn et al.’s image labeling, and the Jaccard coefficient are all ideas that graduate students should see. This paper could therefore be the starting point for a nice seminar course.

These ideas should find widespread adoption in the future, if it can be shown that such a system generates alt-text at a level of quality comparable to an expert. The pressures on colleges and textbook publishers to make STEM education accessible are intense, and this approach has the potential to solve one of the major barriers to doing so.