Trailblazer: Enabling Blind Users to Blaze Trails Through the Web
Jeffrey P. Bigham, Tessa Lau, Jeffrey Nichols · 2009 · IUI '09: Proceedings of the 14th International Conference on Intelligent User Interfaces · doi:10.1145/1502650.1502677
Summary
This paper introduces TrailBlazer, a system designed to help blind web users complete tasks more efficiently by providing step-by-step guidance through web pages. The core problem it addresses is that blind users relying on screen readers must often perform time-consuming linear searches through page content to find specific interactive elements or information — a process that sighted users accomplish almost instantly through visual scanning. TrailBlazer builds on CoScripter, an existing repository of human-created scripts that describe web-based tasks in natural language steps (e.g., "click the Submit button", "enter text into the Search textbox"). The system provides three key contributions: an accessible guide interface designed specifically for screen readers and keyboard-only access; a "clip" command that directs users to specific regions of content on a page identified by heading or contained text; and dynamic script generation that creates new task scripts on-the-fly from a brief user-provided task description combined with the existing script repository. The interface integrates directly into web pages via the DOM, with instruction bubbles placed so that screen readers encounter them in a logical order immediately before the relevant interactive element. Playback controls use keyboard shortcuts (e.g., ALT+S to advance) and are represented as buttons for discoverability. A formative study with 5 blind participants and an evaluation across 15 tasks validated the approach.
Key findings
In the formative evaluation with 5 experienced blind screen reader users (averaging 15 years of computer experience and 11.8 years of web use), participants were enthusiastic about TrailBlazer, with one stating it was "exactly what most blind users would like." Participants uniformly agreed that finding information on complex web pages is time-consuming even with extensive screen reader experience. However, they identified a critical limitation: TrailBlazer was only useful when a pre-existing script was available, and they wanted support for tasks and sites without scripts. This feedback motivated the dynamic script generation feature. In the technical evaluation of suggestion accuracy across 15 popular tasks (102 total steps, averaging 6.8 steps per script), TrailBlazer correctly predicted the next action as its top suggestion 41.4% of the time, and the correct action appeared within the top 5 suggestions 75.9% of the time. Given that the average number of possible actions per step was 41.8 (chance probability of only 2.3%), this represents a substantial reduction in the search space blind users face. The system uses a Naive Bayes classifier trained on six features derived from the task description, script repository, and user action history to rank suggestions.
Relevance
TrailBlazer represents an important early exploration of using intelligent interfaces to reduce the cognitive and temporal burden that blind users face when navigating the web. The fundamental insight — that blind users spend disproportionate time searching for interactive elements that sighted users locate instantly — remains highly relevant today, even as screen readers and web standards have improved. The approach of leveraging crowdsourced procedural knowledge to guide users through unfamiliar websites anticipates modern concepts like browser automation and AI-powered assistants. For accessibility practitioners, this work highlights that making individual page elements accessible (proper labels, headings, landmarks) is necessary but not sufficient; users also need efficient pathways through multi-step tasks. The clip command concept — directing users to specific content regions — foreshadows the importance of ARIA landmarks and skip navigation patterns. While the specific technology is dated, the paper offers lasting lessons about designing non-visual task completion systems and the value of combining human knowledge with machine learning for accessibility.
Tags: screen readers · blind users · web accessibility · non-visual interfaces · programming by demonstration · task guidance · intelligent user interfaces