Avaliação de textos alternativos de imagens na web por meio de comparação entre entropias

Abstract

Accessibility is becoming increasingly more important on the internet, allowing its users to access the contents of webpages with minimum difficulties. One way of providing accessibility for people with visual impairment is via alternative texts associated to images on the internet to describe their content and functionality for screenreader users. It's important, however complicated, to guarantee an alternative text is appropriate for an image and its context to avoid problems in content comprehension. A tool for automatic evaluation of alternative texts incorporated in automated accessibility evaluators would assist web developer even more on guaranteeing the accessibility of their webpages, being one example the tool proposed by Bigham. There are articles that indicate the use of entropies for performing computational analysis on images and texts, which makes them a viable strategy to evaluate alternative texts for their respective images. On the first step of this work, methods for entropy calculation of images and their alternative texts were studied and chosen, being these the entropies of Shannon, Tsallis and Renyi. A collection of images and alternative texts was performed on random Wikipedia pages, and then these images and alternative texts were evaluated by experts via the Likert scale. On the second step of this work, the images with the best evaluations for their alternative texts and their entropies were calculated, and then their correlation was tested via Spearman's equation, which result in a possible rejection of the inexistance of a negative correlation between entropies of images and their alternative texts. However, because of contradictions when the tests were repeated with the images separated in categories, it is probable that more tests have to be made before confirming a correlation between entropies, or that such correlation using the techniques in this work might not exist.