Adapting Online Customer Reviews for Blind Users: A Case Study of Restaurant Reviews
Mohan Sunkara, Akshay Kolgar Nayak, Sandeep Kalari, Yash Prakash, Sampath Jayarathna, Hae-Na Lee, Vikas Ashok · 2025 · Proceedings of the 22nd International Web for All Conference (W4A) · doi:10.1145/3744257.3744276
Summary
This paper addresses the significant usability challenges blind screen reader users face when navigating online customer reviews, using restaurant reviews on Google Maps as a case study. The research proceeds in two phases. First, an interview study with 30 blind screen reader users identified key pain points: listening fatigue after only a few reviews, frustration with repetitive and redundant content, difficulty finding specific information, and outdated reviews that no longer reflect current offerings. Most participants expressed a strong desire for thematic organization of reviews by aspects (food quality, ambiance, hygiene, customer service, pricing) with sentiment-based segregation into positives and negatives. Based on these findings, the authors developed QuickCue, a Chrome browser extension that augments the existing Google Maps review interface with an alternative, screen-reader-friendly presentation. QuickCue uses GPT-4 to perform two core operations: joint classification of reviews into aspect-sentiment pairs using an adapted Clue and Reasoning Prompting (CARP) strategy with few-shot examples, and focused summarization of review subsets using Directional Stimulus Prompting (DSP) to generate concise, aspect-specific summaries. The interface presents content in a three-level accessible accordion hierarchy: aspects at the top level, positive and negative summaries at the second level, and original reviews at the third level, all navigable via standard screen reader keyboard shortcuts.
Key findings
In the interview study, 26 of 30 participants reported struggling with large volumes of reviews, with more than half (17) saying they often stopped after just a few reviews due to listening fatigue. All 30 participants mentioned encountering outdated reviews, with 17 preferring reviews from only the past year. Nearly all (28) wanted aspect-based organization, and two-thirds (19) desired sentiment-based segregation. The joint classifier achieved 0.80 precision, 0.82 recall, and 0.81 F1-score with few-shot prompting, substantially outperforming zero-shot (0.62 F1). In the user study with 10 blind participants, QuickCue achieved a mean SUS score of 81.25 compared to 63.25 for the default screen reader condition, a statistically significant improvement (F = 45.03, p = 2.72 x 10^-6). NASA-TLX workload scores dropped from 62.09 to 38.37 with QuickCue (F = 99.27, p = 9.45 x 10^-9), with the Effort and Temporal Demand sub-scales showing the largest improvements. Seven of 10 participants attributed their positive experience to QuickCue's simplistic design and familiar keyboard navigation, and several participants inquired about extending the tool to other e-commerce platforms.
Relevance
This research highlights a largely overlooked accessibility problem: while much accessibility work focuses on making web content perceivable and operable, the information architecture of user-generated content like reviews creates distinct usability barriers for screen reader users. The linear, sequential nature of screen reader interaction means that content designed for visual skimming becomes tedious and inefficient to consume auditorily. QuickCue demonstrates how LLMs can serve as an intermediary layer to restructure and summarize content into more accessible presentations without modifying the underlying platform. For practitioners, the study provides concrete evidence that aspect-based organization with sentiment separation significantly reduces cognitive load for blind users. The approach is generalizable beyond restaurant reviews to any domain with large volumes of user-generated content, including product reviews, discussion forums, and classified advertisements. The use of ARIA attributes and tab-index for accordion navigation offers a practical template for building accessible hierarchical interfaces.
Tags: screen readers · blind users · online reviews · LLM accessibility · content summarization · sentiment analysis · web accessibility
Standards referenced: WCAG 2.0 · ARIA