Automatically generated headers as text-skimming mechanisms for blind users using screen reading software in unmarked web-based texts

Abstract

The evolution in information access caused by the Internet has expanded access to information for everyone. Consequently, the availability of information sources and teaching content on the Web has broadened access to knowledge. Still, people with disabilities face accessibility barriers. In this work, we focus on people with visual impairment and their problems when reading long documents. Even with assistive technologies such as screen readers, consulting the Web when seeking information can be exhausting for this group. They lack options for speed reading or skimming large amounts of text since these tools read texts aloud in sequence. The commands available in screen readers through shortcuts helped users create strategies when searching for information, such as increased reading speed and using webpage markup to navigate (e.g. headers or paragraphs). However, they often come at the cost of the cognitive load caused by the attention needed to understand and remember the information at high speeds. Navigating inside texts also proved to be a problem. Studies have shown that a common complaint of visually impaired people is having to create a text map to re-find information mentally. To counter these problems, we have conducted a systematic mapping to gather previous work that aimed to help screen reader users when using the Web and categorize them based on their proposed approach. The methods encountered were Content Filtering, Text Reduction, Navigation, Concurrent Speech, Auditory Overview and Recommendation Systems. Based on this analysis, we observed a lack of topicalization techniques and their effects on navigation for blind people. Thus, the goal of this work was to propose an algorithm to generate headers aiming to help users in information-seeking tasks automatically. The algorithm was divided into two tasks: segmenting a document into topic segments and labelling a text segment. We adapted the C99 segmenting algorithm to use BERT and improved error rates for long texts. Then, the study followed with the implementation of a labelling algorithm based on keywords, and labels are made of words from the text segment that were ranked according to repetition. We conducted a user study with 8 participants and a prototype composed of preprocessed texts 720-1131 words long to test the algorithm. Users had to answer questions based on the information in these texts for comparison in two scenarios: one with automatically generated headers and the other without. We measured the time taken in each text and the cognitive load participants perceived while completing it. A post-test interview was also conducted to gather feedback. Our analysis could not confirm our hypothesis significantly due to a small volunteer sample, but interviews indicated users benefited from the proposed tool. Either by helping navigate inside the text, or re-finding information, participants agreed they would like to have this tool in their screen readers. With this work, we provide design implications and alternatives to implement a plugin for the screen reader.