cs.HCAug 26, 2026
SaveQ&A or Document-Based? The Effects of Interface Type on How Screen Reader Users Access Interconnected Documents
Organizations: Department of Computer Science University of Victoria Victoria, British Columbia, Canada · School of Computing and Information Systems Singapore Management University Singapore, Singapore · Lee Kong Chian School of Business Singapore Management University Singapore, Singapore
Abstract
Blind and low-vision (BLV) users are increasingly engaging with large language model (LLM) interfaces to access documents, but it is unclear how such systems support or hinder their ability to build interconnected knowledge. To examine this gap, we compared a Question-Answer Interface (QAI) that supports open-ended conversational inquiry, with a Document Interface (DI) based mostly on traditional structured text document navigation. We recruited 16 BLV screen reader users where they used both interfaces to explore two fictional worlds. Data from interaction logs, concept maps, decision-based tasks, and semi-structured interviews provide comparative insights into how interface design supports knowledge construction. Findings show that participants visited more distinct documents with the DI and formed larger and more correct mental models with the DI than with the QAI. They were also more able to apply knowledge they had gained. Simultaneously, many still preferred the QAI and often estimated that they had explored more, formed better mental models and applied their models better when acquiring the information with the QAI, despite this not being the case. Our analysis suggests possible interface design reasons for these differences and highlights some of the risks introduced by using question-answer interfaces to access information spaces.
Figures & tables
Figure 1. Overview of the study procedure and measurements. Participants completed all phases for each of the DI and QAI conditions in a counterbalanced order. At the end of the second session in Phase 4, participants completed an additional set of interview questions comparing the interfaces. Horizontal process map of the study. The top row shows four phases: Exploration, Verbal Elicitation and Concept Mapping, Decision-Based Task, and Interface Reflection and Comparison. The middle row lists the activity and duration for each phase. The bottom row summarizes the main outcomes, with curved connectors linking concept listing and knowledge application to the shared insight on organization, integration, and application of knowledge.
Figure 2. Experimental setup for the DI condition showing the tactile overlay on top of the tablet (left) on a non-slip mat, with the laptop showing the Dominion interface (right). Setup for the DI condition. On the left, a tablet with a tactile overlay is placed on a non-slip mat next to a laptop, on the right, displaying the Dominion interface.
Figure 3. An example of one of the documents in the DI interface for the world of Solana. A sample document in the DI trial. The website page consists of headers, paragraphs, an overview, and a large main content area with a main heading called 'The Doctrines of Solana.' There re also hyperlinks, providing additional references within the document.
Figure 4. Examples of tactile overlays used in the study: a map (left), a diagram of relationships (center), and a set of icons (right). Tactile overlays. It shows three types of tactile overlays arranged horizontally. The left one represents a map, the center shows a node-link diagram, and the right presents a pyramid of symbols.
Figure 5. The QAI interface for the world of Dominion showing a user query and a structured tabular response. The QAI interface with the Dominion guide. A user asks, "Who are the leaders in Dominion?" with the system responding with a table listing names, seats, and tiers of ten members. The interface contains a heading and a short description of the Dominion.
| ID | Age | Visual ability | Occupation | GenAI tools used | GenAI usage |
|---|---|---|---|---|---|
| 1 | 40–44 | Totally blind | Program coordinator | ChatGPT, Copilot, Be My AI | Daily |
| 2 | 19–24 | Totally blind | Voiceover artist | ChatGPT, Copilot, Gemini, Be My AI | Daily |
| 3 | 50–54 | Low vision | Unemployed | ChatGPT, Be My AI | Daily |
| 4 | 55–59 | Totally blind | Administrative work | ChatGPT, Gemini, Be My AI | 2–3 times a week |
| 5 | 55–59 | Totally blind | Retired | Be My AI | Daily |
| 6 | 25–29 | Totally blind | Unemployed | ChatGPT, Be My AI | Daily |
Table 1. Self-reported demographics of the study participants.
| Overview Predetermined | Overview Self-Defined | Depth Predetermined | Depth Self-Defined | Fact Finding | |
|---|---|---|---|---|---|
| DI | Easy | Hard | Easy | Hard | Hard |
| QAI | Hard | Easy | Hard | Easy | Easy |
Table 2. Summary of the difficulty of using the DI and QAI across the exploration subpatterns, based on qualitative data. Qualitative summary of which interface participants found easier for each exploration intent. The table compares DI and QAI across overview and depth tasks under two conditions: predetermined and self-defined. The table also includes fact finding. Green cells mark the interface reated easier for that subpattern; red cells mark the interface rated harder.
Figure 6. Exploration metrics (RQ1—Phase 1). DI is red, and QAI blue. Numbers above represent the averages. Error bars are standard error. Each black dot is a session. Violin plots comparing DI and QAI on five measures. Unique document visits averaged 11.44 for DI and 9.31 for QAI, p = 0.03. Document transitions averaged 16.31 for DI and 16.81 for QAI, p = 0.85. Document visits averaged 17.31 for DI and 17.81 for QAI, p = 0.85. Unique document transitions averaged 9.56 for DI and 12.63 for QAI, p = 0.06. Transitions per document averaged 1.49 for DI and 1.77 for QAI, p = 0.33.
Figure 7. Exploration graph metrics (RQ1—Phase 1). See caption of Figure 6 for details. Violin plots comparing DI and QAI on four graph measures. Density averaged 0.13 for DI and 0.21 for QAI, p = 0.00. Directed Diameter averaged 7.75 for DI and 4.94 for QAI, p = 0.00. Average Shortest Path Length averaged 3.37 for DI and 2.44 for QAI, p = 0.00. Degree Variance averaged 2.40 for DI and 2.77 for QAI, p = 0.69.
Figure 8. Participant 1’s mental model graphs from the concept mapping. DI (left) and QAI (right). Nodes represent elicited topics (red) and subtopics (yellow). Edges represent explicitly described relationships. On the left, the DI-Solana graph shows interconnected nodes arranged around central topics. There are 11 topics and 9 subtopics. On the right, the QAI-Dominion graph shows 6 topics and 16 subtopics with a more centralized structure with hub nodes connecting to many others. Nodes represent topics or subtopics, and arrows represent directed relationships between them.
Figure 9. Concept map metrics (RQ2—Phase 2). See caption of Figure 6 for details. Violin plots showing five metrics for DI and QAI. Topics averaged 9.44 for DI and 5.63 for QAI, p = 0.00. Subtopics averaged 9.63 for DI and 8.69 for QAI, p = 0.55. Combined topics and subtopics averaged 19.06 for DI and 14.31 for QAI, p = 0.02. Connections averaged 20.69 for DI and 14.94 for QAI, p = 0.01. Connections per topic averaged 1.09 for DI and 1.03 for QAI, p = 0.27
Figure 10. Concept map graph metrics (RQ2—Phase 2). See caption of Figure 6 for details. Violin plots showing four metrics for DI and QAI. Density averaged 0.07 for DI and 0.09 for QAI, p = 0.01. Directed diameter averaged 4.50 for DI and 3.44 for QAI, p = 0.02. Average shortest path length averaged 1.95 for DI and 1.68 for QAI, p = 0.03. Degree variance averaged 2.47 for DI and 1.70 for QAI, p = 0.01.
Figure 11. Decision-based metrics (RQ2—Phase 3). See caption of Figure 6 for details. Violin plots showing five metrics for DI and QAI. Topics averaged 2.44 for DI and 1.38 for QAI, p = 0.02. Subtopics averaged 1.25 for DI and 0.31 for QAI, p = 0.03. Combined topics and subtopics averaged 3.69 for DI and 1.69 for QAI, p = 0.00. Connections averaged 4.19 for DI and 2.50 for QAI, p = 0.09. Connections per topic averaged 1.07 for DI and 0.96 for QAI, p = 0.72
Figure 12. Concept map and Decision-based error metrics (RQ2—Phase 3). See caption of Figure 6 for details. Violin plots comparing DI and QAI across four error metrics. Concept map errors averaged 3.31 for DI and 6.00 for QAI (p = 0.00), while concept map error rates averaged 0.12 for DI and 0.26 for QAI (p = 0.00). Decision-based errors averaged 0.25 for DI and 0.81 for QAI (p = 0.01), and decision-based error rates averaged 0.03 for DI and 0.29 for QAI (p = 0.01).
| Perception | Measurement | Match | |||||
| DI | QAI | DI | QAI | Equal | DI | QAI | |
| Explore more | 10 | 6 | 11 | 4 | 1 | 6 | 2 |
| Learn more | 6 | 10 | 11 | 5 | 0 | 5 | 4 |
| Answer Confidence | 7 | 9 | 12 | 1 | 3 | 5 | 0 |
| Overall Preference | 7 | 8 | N/A | N/A | N/A | N/A | N/A |
Table 3. Comparison of subjective judgments and measured outcomes for DI and QAI. Perception shows how many participants said each interface helped them explore more (RQ1), learn more and answer with greater confidence (RQ2). Measurement shows how many participants actually performed better with each interface. Overall preference is subjective only. Comparison of participant judgments and measured outcomes for DI and QAI. Perception shows reported advantages, Measurement shows metric-based advantages, and Match shows agreement between the two. Overall preference is subjective only.