Accessibility is an indispensable aspect in digital learning. Equitable and proper access ensures all users, including those with disabilities, benefit from content. According to the World Health Organization (WHO), approximately 2.2 billion people today have vision impairment.
To uphold digital accessibility, there are laws like the Americans with Disabilities Act (ADA) and to comply with ADA, the Web Content Accessibility Guidelines (WCAG) have to be followed. These standards ensure that the web content is safe and easy to access. For example, avoiding elements that flash more than three times, as they may induce seizures; Ensuring proper contrasts of textual and background colors for those with vision impairments.
Image
Alt text: Scanning electron micrograph of a blue substrate on which sit many rod-shaped Escherichia coli bacteria, colored light orange.
Generating alt texts effectively presents several challenges that arise from human limitations and the current state of automated tools. About 90.9% of blind or visually impaired people rely on screen readers. Inaccurate alt text can lead to misinterpretations and decrease credibility and user experience. Therefore, it is crucial to identify challenges to arrive at effective techniques.
What Are the Challenges That Humans Face?
- Creating alt texts manually when there are larger volumes of images becomes tedious. Each image description must be accurate and unique. Hence, it can also become overwhelming and impractical to manage deadlines.
- Human errors are inevitable. Basic typos or chances of misinterpreting the image are a high possibility. Multiple checks are required leading to additional resources.
- Certain images are challenging to describe due to its complex nature. It can lead to oversimplified or overly detailed alt texts. Oversimplified alt texts fail to describe the images. Overly detailed ones are difficult to follow.
For instance, let us look at the life cycle of the malaria parasite below:
Poor Alt Text
- Oversimplified: A diagram showing the life cycle of the malaria parasite. The cycle starts with transmission of parasite cells to a human via a mosquito bite and ends with the cells of the parasite changing shape in preparation to infect the human host.
- Overdetailed: A diagram showing the life cycle of the malaria parasite. The cycle starts with transmission of parasite cells to a human via a mosquito bite. The parasite is shown infecting liver cells. These cells then rupture, allowing entrance of the parasite cells into the bloodstream. Next the red blood cells are shown being infected by the parasite. This leads to a cycle of asexual reproduction of the parasite within the red blood cells; as the parasite reproduces, the red blood cells rupture, releasing more parasites into the bloodstream, which infect other red blood cells and continue to reproduce. Next the parasite is transmitted to a mosquito when the mosquito bites an infected person. Next the parasite undergoes sexual reproduction inside the mosquito. Next growth of the parasite in the mosquito occurs; cells are shown multiplying. Finally, the cells of the parasite change shape in preparation to infect the human host. The cycle begins again when the parasite is transmitted to a human via a mosquito bite.
Good Alt Text
A diagram showing the life cycle of the malaria parasite. The cycle starts with transmission of parasite cells to a human via a mosquito bite. Next the parasite is shown infecting liver cells. Next the liver cells are shown rupturing, allowing entrance of the parasite cells into the bloodstream. Next the red blood cells are shown being infected by the parasite. This leads to a cycle of asexual reproduction of the parasite within the red blood cells. The parasite is transmitted to a mosquito when the mosquito bites an infected person. Then the parasite undergoes sexual reproduction inside the mosquito. The parasite multiplies and the cells change shape in preparation to infect the human host. The cycle begins again when the parasite is transmitted to a human via a mosquito bite.
Comparison of Current Automated Tools
Here’s a high-level comparison of two of the competitive alt text automation tools available today:
Challenges of Using Existing AI tools
- Automated tools generate random and irrelevant image descriptions. These tools rely on algorithms. They do not understand the context of the image, leading to vague information. This makes it difficult for the visually impaired to grasp the meaning of the image.
- It also misinterprets the image, leading to incorrect descriptions.
- Most of the current AI tools lack the ability to perceive details. It leaves out essential information that is required to imagine the image correctly.
There are effective strategies and technologies that can help us bridge the gaps we face during manual and automated generation.
Making Alt Texts Effective
- Avoid using “Image of”. Screen readers already indicate that an image is present. Instead mention a chart, an illustration, the life cycle, etc.
- Avoid repeating the image caption as alt text.
- Be specific. Do not leave out important information.
- Include text provided in the image.
- Keep it concise. Not more than 100–150 words.
- Spell them out clearly so that the screen readers will read them exactly how you expect them to be.
- Use Unicode characters. Not all screen readers support special characters and symbols.
Techniques for Automating Alt Text Generation
We can significantly enhance the quality and efficiency of alt text generation through these techniques to generate contextually relevant, accurate, and comprehensive descriptions.
Six Steps to Automate Your Alt Texts
- Identify image repository.
- Run and use BLIP (a machine learning model) to generate text descriptions.
- Run a script using Python or Node.js to automate step 1 and 2.
- Manually review the text generated to ensure they are accurate and relevant.
- Refine the text using a natural language processing tool like ChatGPT to improve context of the description.
- Collect feedback from reviewers and users to identify issues in the generated alt text.
- Update AI model and scripts to improve accuracy and relevance.
Human Intervention Is A Must!
These tools, coupled with the usage of AI today, certainly offer great upside and potential. While the future is exciting, there are still some questions around the efficacy of automation in alt text generation. We, at EduQual, recommend a human + tool approach to generating relevant and accurate alt texts.
While the tool does the heavy lifting production, our Subject Matter Experts (SMEs) validate alt texts for accuracy, relevance, and cultural insensitivity. This brings down the production time by over 50% while maintaining 99%+ accuracy levels. This is why we can’t overlook the importance of human intervention in this process.
Sources
1.https://www.researchgate.net/publication/363920555_A_Dataset_of_Alt_Texts_from_HCI_Publications_Analyses_and_Uses_Towards_Producing_More_Descriptive_Alt_Texts_of_Data_Visualizations_in_Scientific_Papers
2.https://proceedings.neurips.cc/paper_files/paper/2019/file/680390c55bbd9ce416d1d69a9ab4760d-Paper.pdf
3.https://lab.co.uk/news-and-insights/automating-alt-text-generation-in-prismic
4.https://pmc.ncbi.nlm.nih.gov/articles/PMC9395872/
Written by Anna Riya R
