Paper Reading #16: Mixture Model based Label Association Techniques for Web Accessibility

Commentary

See what I have to say about Wesley's and Miguel's work.

References

Islam, M. A., Borodin, Y., Ramakrishnan, I. V. (2010). Mixture Model based Label Association Techniques for Web Accessibility. Proceeding of the Acm conference on user interface software and technology. New York: http://www.acm.org/uist/uist2010/.

Article Summary

Islam et al. present a system they have created that extends the functionality of an assistive technology called screen reading. This technology utilizes text-to-speech to allow blind users to navigate websites by reading the content on webpages and descriptions of elements to them. A major impediment to the proper functioning of screen readers is the omission of labels for page elements and alternative text for images. Without proper labels, form elements can be misrepresented or not denoted at all. Without alternative text for images, transaction functionality for most websites is completely lost, as transactional dialogs are usually controlled through images, e.g. an "Add to cart" or "Checkout now" button:

Taken from the above referenced paper.

Even properly labeled items are sometimes not handled properly by the screen reader by virtue of the ambiguity of the HTML Document Object Model (DOM), e.g. labels for elements and the elements themselves being contained in different HTML table rows:

Taken from the above referenced paper.

The authors implemented a finite mixture model (FMM) to create contexts to which HTML elements and possible labels belong. Using these contexts, the FMM can also create labels for unlabeled elements with some accuracy and more correctly interpret labels for ambiguous objects. In evaluating their system, the authors observed a 76% success rate of correctly applying labels to their elements without prior training by their FMM and a 95% success rate with prior training when all elements were explicitly labeled. On a testing set without any labels, their FMM achieved an 81% success rate. In evaluating their system through a user study with two blind users who were proficient with screen reading technology, both blind users agreed that the FMM made interacting with webpages easier for them.

Discussion

Wow. I personally feel that this is some of the most incredible research I've discovered yet. The authors have created a system that solves a very practical problem, and solves it well. The idea of creating contexts from which to infer labels was ingenious, as was the systematic approach to evaluating documents geometrically. I have to commend them: the system they created seems to be entirely robust. In addition, while this work may have implications outside of catering to users who need assistive technology, I feel that there was at least some measure of philanthropic drive behind the project, purposefully or not. I approve of this work.