Developing a Corpus of Hierarchically Classified STEM Images for Accessibility Purposes

Abstract

Even though considerable efforts have been made to provide effective image descriptions for digital accessibility, a large portion of STEM images, especially complex STEM images, nowadays remains inaccessible to people with visual disability. The quality of alt text is much more critical for university STEM textbooks as image descriptions must be accurate and detailed but not tire out the reader. This work aims to develop a large corpus of hierarchically classified STEM images from university textbooks which later will be used for developing appropriate guidelines for meaningful non-automatic high-quality alt-text image descriptions with the purpose of accessibility in mind. We present first our approach for the creation of the corpus with the STEM images. Our corpus at the current stage includes 8.859 STEM images from 82 textbooks in the domains of Mathematics, Biology, Computer Science, Chemistry, Physics and Geology. Then, we describe the methodology we followed for the classification of the images in the corpus and in particular, the way for the creation of the five categories and twenty-four subcategories, as well as the manner of the assignment of images to categories.