QuickQue: Enabling Quick Access to Information in User Reviews for Screen Reader Users
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 2025) · doi:10.1145/3744257.3744279
Summary
This paper presents QuickCue, a Google Chrome browser extension that helps blind screen reader users efficiently access online customer reviews by using LLM-powered aspect and sentiment classification to organize and summarize review content. The current experience of reading online reviews with a screen reader is highly inefficient: reviews are presented as a linear list that must be navigated sequentially, containing redundant information across multiple reviews about the same topics. Using restaurant reviews on Google Maps as a case study, QuickCue addresses this by performing two core operations via GPT-4. First, joint classification categorizes each review by its applicable aspect-sentiment pairs (e.g., food quality/positive, hygiene/negative), recognizing that a single review can span multiple aspects with different sentiments. This uses the CARP (clue and reasoning) prompting strategy. Second, aspect-focused summarization generates concise summaries for each aspect-sentiment group using directed stimulus prompting (DSP). The accessible interface presents information as an HTML accordion with five aspect dropdown buttons (food, hygiene, pricing, customer service, ambiance), each expanding to show positive and negative summaries, with access to the original classified reviews beneath. The interface uses ARIA attributes and tab-index for screen reader navigation.
Key findings
A user study with 10 blind participants showed significant improvements in usability and task workload when using QuickCue compared to the standard Google Maps screen reader experience. The tool reduced information redundancy by grouping reviews thematically and providing concise summaries, allowing users to quickly get an overview of positive and negative aspects without sifting through numerous individual reviews. The hierarchical interface structure—aspects, then sentiment summaries, then original reviews—enabled users to drill down to their desired level of detail using familiar keyboard navigation (TAB and SHIFT+TAB). The system extracted reviews from Google Maps using pre-defined XPath information to identify review DOM nodes, processed text using NLTK and spaCy for noise removal, and integrated with GPT-4 via the LangChain framework. All data exchanges used JSON format for consistency between QuickCue modules.
Relevance
QuickCue addresses a practical and overlooked accessibility gap: the difficulty blind users face when trying to make informed decisions based on online reviews. While reviews are a critical resource for consumer decisions—choosing restaurants, products, services—the linear, redundant nature of review lists creates a disproportionate burden for screen reader users who cannot visually scan and skim content. The aspect-based organization and summarization approach transforms an inherently visual browsing experience into one that works well with the sequential nature of screen reader interaction. For practitioners, the work demonstrates how LLMs can be applied not just to generate accessible content but to restructure existing inaccessible content into accessible formats. The browser extension approach is particularly practical as it enhances existing platforms without requiring changes from the platform providers, giving users agency to improve their own browsing experience.
Tags: screen readers · blind users · online reviews · large language models · browser extension · information accessibility · sentiment analysis
Standards referenced: ARIA