Screen-reader software enables blind users to access large segments of electronic content, particularly if accessibility standards are followed. Unfortunately, this is not true for much of the content written in physics, mathematics, and other STEM-disciplines, due to the strong reliance on mathematical symbols and expressions, which screen-reader software generally fails to process correctly. A large portion of such content is based on source documents written in LaTeX, which are rendered to PDF or HTML for online distribution. Unfortunately, the resulting PDF documents are essentially inaccessible, and the HTML documents greatly vary in accessibility, since their rendering using standard tools is cumbersome at best. The paper explores the possibility of generating standards-compliant, accessible HTML from LaTeX sources using Large Language Models. It is found that the resulting documents are highly accessible, with possible complications occurring when the artificial intelligence tool starts to interpret the content.
翻译:屏幕阅读软件使得盲人用户能够访问大部分电子内容,尤其是在遵循无障碍标准的情况下。遗憾的是,对于物理学、数学及其他STEM学科中的大量内容而言,由于这些内容强烈依赖数学符号和表达式(屏幕阅读软件通常无法正确处理这些内容),这一说法并不成立。此类内容中有很大一部分基于用 LaTeX 编写的源文档,这些文档会被转换为 PDF 或 HTML 格式以便在线分发。然而,最终生成的 PDF 文档基本上无法访问,而 HTML 文档的可访问性也参差不齐,因为使用标准工具对其进行渲染充其量也相当繁琐。本文探索了利用大型语言模型从 LaTeX 源生成符合标准且可访问的 HTML 的可能性。研究发现,生成的文档具有高度的可访问性,但当人工智能工具开始解释内容时,可能会出现一些复杂情况。