In a data-driven society, being able to examine data on one’s own terms is crucial for various aspects of well-being. However, current data exploration paradigms, such as Exploratory Data Analysis (EDA), heavily rely on visualizations to unveil patterns and insights. This visual-centric approach poses significant challenges for blind and low-vision (BLV) individuals. To address this gap, we built a prototype that supports non-visual data exploration and conducted an observational user study involving 18 BLV participants. Participants were asked to conduct various analytical tasks, as well as free exploration of provided datasets. The study findings provide insights into the factors influencing inefficient data exploration and characterizations of BLV participants’ analytical behaviors. We conclude by highlighting future avenues of research for the design of data exploration tools for BLV users.

In recent years, there has been a growing interest in enhancing the accessibility of visualizations for people with visual impairments. While much of the research has focused on improving accessibility for screen reader users, the specific needs of people with remaining vision (i.e., low-vision individuals) have been largely unaddressed. To bridge this gap, we conducted a qualitative study that provides insights into how low-vision individuals experience visualizations. We found that participants utilized various strategies to examine visualizations using the screen magnifiers and also observed that the default zoom level participants use for general purposes may not be optimal for reading visualizations. We identified that participants relied on their prior knowledge and memory to minimize the traversing cost when examining visualization. Based on the findings, we motivate a personalized tool to accommodate varying visual conditions of low-vision individuals and derive the design goals and features of the tool.

Data-driven news articles are widely used to communicate societal phenomena with concrete evidence. These articles are often accompanied by a visualization, helping readers to contextualize content. However, blind and low vision (BLV) individuals have limited access to visualizations, hindering a deep understanding of data. We explore the possibility of dynamically generating data facts (texts describing data patterns in a chart) for BLV individuals based on their preferences to aid the reading of such articles. We conduct a formative study to understand how they perceive system-generated data facts and the factors influencing their preferences. The results indicate the preferences are highly varied among individuals, and a simple preference elicitation alone induces noise. Based on the findings, we developed a method to personalize the data facts generation using an active learning approach. The evaluation studies demonstrate that our model converges effectively and provides more preferable sets of data facts than the baseline.

The data table is a basic but versatile representation to communicate data. From government reports to bank statements, tables effectively carry essential data-driven information by visually organizing data using rows, columns, and other arrangements (e.g., merged cells). However, many tables online neglect the accessibility requirements for people who rely on screen readers, such as people who are blind or have low vision (BLV). First, we consolidated guidelines to understand what makes a table inaccessible for BLV people. We conducted an interview study to understand the importance of tables and identify further design requirements for an accessible table. We built a tool that automatically detects HTML formatted tables online and transforms them into accessible tables. Our evaluative study demonstrates how our tool can help participants understand the table’s structure and layout and support smooth navigation when the table is large and complex.

The task of answering questions about images has garnered attention as a practical service for assisting populations with visual impairments as well as a visual Turing test for the artificial intelligence community. Our first aim is to identify the common vision skills needed for both scenarios. To do so, we analyze the need for four vision skills--object recognition, text recognition, color recognition, and counting--on over 27,000 visual questions from two datasets representing both scenarios. We next quantify the difficulty of these skills for both humans and computers on both datasets. Finally, we propose a novel task of predicting what vision skills are needed to answer a question about an image. Our results reveal (mis)matches between aims of real users of such services and the focus of the AI community. We conclude with a discussion about future directions for addressing the visual question answering task.