Machine learning for accessible web navigation

Abstract

This research looks at the application of Machine Learning to Web Accessibility. It considers how Machine Learning (ML) can be used to help make the processes of Web Navigation more accessible in line with Web Content Accessibility Guideline (WCAG) 2.4 Navigable, which demands that ways be provided "to help users navigate, find content, and determine where they are." ML techniques such as reinforcement learning have been applied to website navigation in diverse ways. These include goal-directed search to answer questions and task-oriented problems such as booking a flight. Related work includes Web Automation and Testing. These techniques typically involve a state space exploration where actions such as clicking links, filling forms, and pressing buttons move the agent to a new state. The search space can be represented as a directed graph whose nodes represent the state. The choice of action involves a reward in reinforcement learning, which is used to learn behavior. Sequences of actions are collected into policies and the objective is to identify the best policy. For many websites, the potential search space is exceptionally large, and these techniques can provide ways of navigating them more efficiently. While the outcomes of these approaches can help to make the processes of search and task completion easier and thereby help accessibility, it is not the primary focus. The question under consideration is whether these techniques can be adapted to improve the website's navigability from an accessibility perspective. The approach seems promising. The state-space can be explored automatically, and the state representation can then be mined using ML techniques for structure, content and other information, which has the potential to improve accessibility Examples of the application of these approaches, which reflect the goals of the success criteria around WCAG Guideline 2.1 would include the generation of good anchor text for links where this is not provided, optimizing pathways to website functionality, and optimizing the processes of querying the website. There are open questions about how state representation can be enhanced to improve the prospects of reaching goal states. For example, can the process of calculating reward exploit features such as Search Engine Optimization information, which is designed to reflect the purpose of the page directly? Indexing techniques from Information science can serve a similar purpose. From an accessibility standpoint, there are exciting directions to explore. For example, what role can accessibility features, e.g., Headings, Anchor Text, and alternative text, play in calculating rewards. By way of illustration, a case study on one of these page enhancement techniques, DocTTTTTQuery, will be presented. This enhances a document by generating potential queries from its content. The system is trained by matching these generated questions against a set of historical questions asked of the site. This knowledge can then be exploited to lead the user more directly to potential answers to queries. Other studies have shown the effectiveness of this approach. Augmenting the page with these queries has shown improved search performance.