Highly Accessible Large Language Model: Designing Inclusive Prompt-Based Interfaces for Knowledge Exploration via LLMs

The increasing adoption of Large Language Models (LLMs) has changed how individuals engage in knowledge exploration activities, providing new ways to explore and consume knowledge through conversational interfaces. However, these interfaces often fail to meet the accessibility needs of Blind and Visually Impaired (BVI) users. In fact, as current chat-based interfaces for LLMs, such as ChatGPT, fail to prioritize accessibility, assistive technologies like screen readers encounter significant difficulties in enabling a more accessible interaction with LLMs. While voice-based interfaces are available, current implementations are often insufficient for BVI users, offering a limited control over navigation and orientation within these systems. To address the accessibility challenges faced by BVI users with conversational interfaces for LLMs, this thesis introduces HALLM (Highly Accessible LLM), a prompt-based prototype system for enhanced and controllable voice-based interaction. In order to develop HALLM, a human-centered approach based on formative studies with BVI users was adopted. HALLM was then evaluated through a series of focus groups that provided valuable insights into its effectiveness and usability, indicating that the prototype effectively addresses significant challenges faced by BVI users. However, although HALLM addressed key issues, additional improvements in its flexibility and integration are needed to develop more inclusive and adaptable voice-based interfaces for LLMs. Despite these challenges, this work proves that conversational voice-based interfaces like HALLM demonstrate significant potential for enhancing knowledge exploration for BVI users.