MatplotAlt: A Python Library for Adding Alt Text to Matplotlib Figures in Computational Notebooks

Area of Science:

• Data Visualization
• Scientific Computing
• Accessibility

Background:

• Matplotlib is a widely used Python library for creating static, animated, and interactive visualizations.
• Generating descriptive alternative text (alt text) for complex data visualizations remains a challenge, impacting accessibility for visually impaired users.
• Existing methods for automated alt text generation often lack accuracy or customization options.

Purpose of the Study:

• To introduce MatplotAlt, an open-source Python package designed to simplify the addition of alt text to Matplotlib figures.
• To enable Jupyter notebook authors to automatically generate and display chart descriptions with minimal code.
• To provide customizable options for alt text generation and display, catering to user preferences and accessibility requirements.

Main Methods:

• Developed MatplotAlt as a Python package integrated with Jupyter notebooks.
• Implemented heuristic-based and large language model (LLM)-based methods for generating alt text.
• Evaluated the accuracy of generated alt text for various Matplotlib figures, including univariate and complex plots.
• Investigated methods to improve LLM accuracy, such as prompting with heuristic-based alt text or parsed data tables.

Main Results:

• MatplotAlt successfully generates accurate long-form descriptions for both simple and complex Matplotlib figures.
• Both heuristic and LLM-based methods within MatplotAlt demonstrate effectiveness in creating descriptive alt text.
• State-of-the-art LLMs exhibit factual inaccuracies when describing charts independently.
• Prompting GPT-4 Turbo with heuristic-based alt text or data tables significantly enhances description accuracy.

Conclusions:

• MatplotAlt offers a practical and efficient solution for enhancing the accessibility of Matplotlib visualizations.
• The package empowers researchers and educators to create more inclusive data narratives within Jupyter environments.
• Combining heuristic approaches with LLMs, particularly through informed prompting, is crucial for improving the factual accuracy of automated chart descriptions.