- The comprehensive paper by [Authors] presents Arboretum, a web system converting Mermaid and Graphviz specifications into accessible forms (tabular, navigable, tactile) enabling students to study and reason about data structures more easily.
- Their Wizard-of-Oz study of three nonvisual diagram prototypes derived five design requirements, emphasizing structural information, separation of facts from explanations, and complementary modalities.
- An evaluation with eight participants found highly accurate responses and a clear preference for tactile graphics, indicating that while no single modality was sufficient alone, multimodal synchronization facilitated better comprehension of data structures.
Overview
"Nonvisual Support for Understanding and Reasoning about Data Structures" (2601.19168) addresses a persistent gap in accessible computer science education: blind and visually impaired (BVI) students encounter data structures primarily through visual diagrams, while standard accommodations such as ALT text preserve visual appearance rather than the computational properties—roles, relationships, ordering, and traversal logic—that make diagrams useful for reasoning. The authors argue that this lossy translation is the core barrier, and that accessibility should instead be grounded in the structural information already present when diagrams are authored in text-based specification languages.
The paper makes three contributions: (1) five design requirements derived from a Wizard-of-Oz study of three nonvisual diagram prototypes; (2) Arboretum, a web-based system that compiles Mermaid and Graphviz DOT specifications of arrays and binary trees into three synchronized accessible representations (tabular, navigable, tactile); and (3) an evaluation with eight BVI participants demonstrating that these representations support comprehension and algorithmic reasoning, including binary search.
Motivating problem
The authors frame the problem through the case of a student receiving ALT text for an array or binary tree. To find the element at index 6, the student must count sequentially through values; to verify the binary search tree (BST) property, they must repeatedly parse prose descriptions of parent–child links, overloading working memory. This experience meets the minimum legal standard under WCAG "complex image" guidance and the 2024 DOJ Title II rule, yet falls short of equitable access because diagrams are treated as images to be described rather than structured representations that can be encoded directly.
The key insight is that text-based diagram specification languages (Mermaid, Graphviz DOT, PlantUML, d2) already capture nodes, edges, and layout rules in machine-readable form. If compilation starts from these specifications rather than from rendered images, accessible multimodal outputs can be generated automatically at near-zero marginal cost to educators—a decisive advantage over hand-produced tactile graphics, which require specialized equipment and per-diagram labor.
Wizard-of-Oz study and design requirements
Before building the system, the authors ran a Wizard-of-Oz study with eight BVI participants (ages 36–67) exploring flowchart representations in three formats: enriched ALT text with LLM-backed question answering, a keyboard-navigable digital graph, and a swell-paper tactile graphic with a digital legend. Affinity analysis yielded five design requirements:
- DR1 — Standardized screen reader navigation: custom directional navigation matched to flowchart semantics confused participants; familiar ARIA-style commands were preferred.
- DR2 — Structural information in tables: tabular layouts aligned with existing screen reader table-navigation expertise and supported random access.
- DR3 — Explicitly encoded relationships: linearized node/edge lists forced participants to infer connectivity.
- DR4 — Separation of structural facts from narrative explanation: mixed "data dump" descriptions obscured what was explanatory versus structural.
- DR5 — Integrated access across complementary modalities: no single representation sufficed; participants spontaneously cross-checked between formats.
These requirements directly shaped Arboretum's architecture and its decision to generate multiple synchronized outputs rather than a single best alternative format.
The Arboretum system
Arboretum is a client-side TypeScript/React application following a three-stage pipeline. In the input stage, educators author diagrams in Mermaid or Graphviz DOT and declare the data structure type (currently array or binary tree). The translation stage parses the specification into an intermediate representation (IR) encoding accessibility-relevant semantics—nodes, edges, roles, ordering—and derives additional properties such as node depth and position. Because all outputs compile from the same IR, any input edit propagates automatically to every modality, which the authors identify as the mechanism for consistency across representations (their answer to DR5).
The output stage generates:
- Tabular: HTML tables making relationships explicit—for trees, columns for value, parent, left/right child, and positional role—satisfying DR2 and DR3.
- Navigable: screen reader-friendly HTML preserving visual layout; arrays render as ARIA lists of index/value pairs, trees as nested ARIA tree structures with expand/collapse navigation, satisfying DR1.
- Tactile: scalable SVGs with Braille-labeled nodes and arrowed directed edges, producible on standard embossers or Swell Form machines found in university assistive technology offices.
The interface offers an editor mode with live synchronized preview for educators and a simplified preview mode for students, developed against WCAG 2.1 AA and tested with JAWS, NVDA, and VoiceOver.
Evaluation with BVI users
Eight BVI participants (ages 37–65; four with some prior CS exposure) completed six tasks in a within-subjects design progressing from array element location through tree relationship identification to binary search. Participants could freely choose among representations after a counterbalanced learning phase. Quantitative analysis used Accuracy of Extracted Information (AEI) plus task timing; qualitative analysis applied open coding (444 initial codes consolidated into 78 focused codes and five themes). One task (T2, order/sorting) was excluded because anomalously fast response times (~2 s), absent strategy descriptions, and lack of correlation with other measures suggested answers from memory.
Performance was strong overall. Element location reached 91.67% accuracy (M = 22.8 s); parent/child identification was perfect (100%, M = 6.9 s); and binary search achieved 100% accuracy across all queries (M = 35.5 s). The hardest task was BST property verification at 62.5% accuracy and the longest mean time (62.5 s), which the authors attribute to the need to simultaneously track values, left/right positioning, and depth. Confidence ratings were high throughout (4.13–4.88 on a 5-point scale).
Representation preference was emphatic. Tactile graphics received uniform ratings of 5.0 across all measures, while tabular and navigable ratings dropped from 4.25 for arrays to 3.38 for binary trees. Of 40 task decisions, participants chose tactile alone 31 times, combined modalities 6 times, and non-tactile modalities only 3 times; seven of eight participants spent over 89% of recorded task time on tactile graphics. Some participants described tactile use as "cheating" because it made tasks so easy—an indication of how much access digital alternatives had previously denied them.
Qualitatively, the results support the paper's central claims. Tactile graphics enabled holistic spatial scanning—for example, verifying the BST property by reading left-to-right value ordering rather than checking node by node ("the details... popped"). Digital representations showed structure-specific weaknesses: the navigable tree's expand/collapse semantics conflicted with spatial arrow-key expectations, and the tabular tree was criticized as behaving "like a 2D array, it doesn't behave like a tree." Multimodal integration emerged as a genuine strategy: participants used tactile graphics to anchor mental models, then verified answers in tables or navigable views, describing the representations as "building blocks" that reinforced one another. Participants also exhibited metacognitive awareness, monitoring whether their representational strategy preserved the conceptual intent of a task (e.g., refusing to treat a tree table as an array). A final theme documented systemic educational barriers: participants recalled lacking accessible materials, relying on self-teaching, and encountering instructors who had no resources for teaching data structures to blind students.
Discussion: design principles and scalability
The authors consolidate their requirements into four principles extending WCAG's POUR framework to structure-first accessibility. Perceivable requires explicit semantic roles and relationships, since a single node occupies multiple dynamic roles (left child, subtree root, leaf candidate) that must be encoded rather than inferred. Operable calls for structure-aligned navigation; the paper argues that WAI-ARIA patterns presuppose hierarchical content and break down for relational structures, suggesting systems should expose multiple traversal logics (e.g., depth-first versus breadth-first). Understandable prioritizes structural reasoning, favoring explicit positional cues even at some cost in navigation speed. Robust reframes cross-modality consistency: divergence arises when each modality interprets a visual source differently, so representations should originate from a shared structural model—the architectural choice Arboretum embodies.
On scalability, the paper emphasizes instructor-facing accessibility. Most prior work supports learners directly but leaves educators without tools to author accessible materials, despite institutional mandates to do so. By accepting specification languages instructors may already use, generating consistent outputs across large teaching teams, and supporting ad-hoc authoring (e.g., constructing a custom tree during office hours), Arboretum treats accessibility as an extension of existing curricular practice rather than an accommodation layered afterward.
Limitations and open questions
The paper is candid about constraints. Input support covers only Mermaid and Graphviz DOT, and array creation is limited to Mermaid; educators unfamiliar with these languages face a learning curve. Screen reader validation covered only JAWS, NVDA, and VoiceOver (macOS), leaving Narrator and mobile screen readers untested. Labels are length-limited (roughly ten characters visually, three in Braille). Tactile media inherits physical limits—sheet size, resolution, production cost—and does not scale to very large structures, while large tables can themselves become cumbersome under screen reader navigation. The evaluation involved small samples (n = 8 in both studies) with descriptive statistics only, focused on introductory structures and tasks, and the remote format limited direct observation of tactile interaction. One participant was excluded for insufficient screen reader familiarity despite screening, illustrating recruitment and screening challenges in this population. Whether the derived principles extend to graphs, linked lists, multidimensional arrays, and more advanced algorithmic reasoning remains an open empirical question the paper explicitly defers to future work.
Conclusion
This paper reframes accessible data structure education around preservation of computational properties rather than description of visual appearance. Its evidence chain—from Wizard-of-Oz-derived requirements, through a compiler-based system producing synchronized tabular, navigable, and tactile outputs from a shared intermediate representation, to a user study showing high accuracy including perfect performance on binary search—supports the claim that structure-first, multimodal generation is both practical for educators and effective for BVI learners. The strong preference for tactile graphics, alongside demonstrated value of cross-modal verification, indicates that no single modality suffices and that synchronization from a common structural model is the appropriate foundation. The work leaves open extension to richer diagram types, broader assistive technology compatibility, and larger-scale studies, but establishes a concrete, reproducible pattern for integrating accessibility into instructional material creation itself.