Segmentação e classificação de ARIA landmarks em aplicações web

Abstract

Digital accessibility is an essential aspect of ensuring the inclusion of people with disabilities in the use of web technologies, enabling a more efficient browsing experience in accordance with the Web Content Accessibility Guidelines (WCAG). Based on this premise, the present work addresses the segmentation and classification of ARIA landmarks in web pages, elements that are critical for digital accessibility, especially for users with visual impairments. Despite the WCAG and the importance of ARIA landmarks, many web pages still fail to implement them correctly due to technical complexity and developers' lack of knowledge. This research proposes the use of supervised machine learning to automate the identification and categorization of landmarks, aiming to improve navigability and compliance with accessibility standards. The methodology involved data collection using a Selenium- based crawler, labeling of elements according to the W3C specifications for ARIA roles and HTML5 tags, and the engineering of numerical and positional features. Three algorithms were trained and evaluated: Decision Tree, Support Vector Machine (SVM), and Multilayer Perceptron (MLP). The results of the comparative analysis indicated that the Decision Tree model achieved the best performance, outperforming the other algorithms and demonstrating to be the most effective approach for landmark classification with the dataset and features used. Thus, for a set of geometric and positional features, rule-based models prove more effective compared to more complex models. Finally, by presenting an automated method that can be integrated into development and testing tools, this work contributes to the field of Computer and Software Engineering, in addition to providing a foundation for the creation of systems that validate and correct web page semantics, thereby promoting a more inclusive and accessible internet.