 Hi, my name is Elisa, and I'm delighted to present this work on characterizing image accessibility on Wikipedia across languages, which is a collaboration between researchers at Stanford and Cuba. Using Wikipedia is designed as an eye-opening experience on many levels, and this experience isn't only metaphorically, but also literally very visual, with many, many images that illustrate the content. On English Wikipedia alone are an estimated 5 million distinct and contentful images. In fact, images are Wikipedia's most common multimedia content and lead to the highest multimedia engagement. Previous work has shown that users especially engage with images and articles about art and less well-known people. So images are a useful and appreciated feature in the Wikipedia experience. But due to their visual nature, they pose an accessibility challenge, especially for blind and low-vision users who have to rely on screen readers to read out the content of a page. For these users, images are only accessible if they have an associated alt description that can be read out. It's well documented that this is a major issue on social media platforms. For instance, on Twitter, only 0.1% of all images have alt text. In this work, we are asking how accessible are Wikipedia's images across languages. We highlight the challenges that come with estimating accessibility coverage and list potentially promising mitigation strategies that can help provide a more inclusive experience. Let's first see what type of image-related texts there are and how well they're covered across languages. We care about all of the texts that relate to an image since they all might potentially aid the accessibility purpose to some degree. We see here the Wikipedia article for banana and are in the section on Cavendish bananas. Below the image in this paragraph is what tends to be called a caption. Here, Cavendish bananas are the main commercial banana cult of art sold in the world market. The caption isn't intended to replace the image, but instead to add contextual information that connects the image to the article content. Another associated text is the alt text, which is part of the HTML signature of the image. And this is what a screen reader will read out if present to make the image accessible. The alt text present for this image is grocery store photo of several bunches of bananas. This is the main text important for accessibility, which we will focus on. The last image based text on Wikipedia is what we call the attribution description. This is a description associated with the image when viewing it on Wikimedia Commons. In contrast to the alt text, this description is independent of any article where the image might appear. It also isn't intended to replace the image, but also frequently contains content about the person who uploaded the image when it was taken, image size specifications, etc. Now, all of these texts relate to the image directly, so let's inspect how often they're present on Wikipedia. This data analysis is performed on the with data set, which contains crawled Wikipedia data from 2020, covering 108 languages. Let's start with attribution descriptions, so the ones that are on Wikicommons. Across languages, almost all images have some attribution description present. This could be however in a language that's different from the language of the article where the image is in. About 50% of images have captions, that is the text appearing below an image. Alt descriptions for accessibility are the least available, covering on average 10% of the images in a given language. And this already points to a universal image accessibility issue on Wikipedia. So, using widths, we can easily extract whether a type of text is present. But this data can only function as upper estimates for how many of those texts are actually useful. Since alt texts are generally not visible to the Wikipedia users and editors, they suffer from severe quality issues. So how can we narrow this down a bit more? One indication of alt description quality in a language is the number of unique descriptions. Informative descriptions can be expected to be quite specific to the image they describe. In most languages, about 80% of all alt descriptions are unique. Noteworthy outliers are German and French with only about 65% unique descriptions. And in fact, when following up on these results, we find that in German, the high duplication rate is due to extensive use of image galleries and articles, where the alt description is defined by the template engine to be the same as the caption. To investigate whether there are other potentially general estimates of the proportion of useful descriptions, we conducted a human subject experiment where we showed participants sample descriptions from English Wikipedia. They rated those using established criteria from prior research and were presented within the available Wikipedia context to imitate the user experience as closely as possible. Unfortunately, we don't find evidence for a generally applicable heuristic to detect a useful from a non useful alt description. That doesn't seem to be a firm minimum length requirement, nor are descriptions that are repetitions from the main text uniformly considered bad. From the 30 descriptions tested only two are consistently rated as good by the annotators. About half of the descriptions are on average rated as rather not good, so less than three out of five on the Likert scale. And these results suggest that even for English, the rate of overall alt text coverage potentially severely overestimates the rate of useful descriptions. But automatically detecting and flagging low quality descriptions will require sophisticated methods guided by empirical research. So what can we do going forward. Firstly, it might be helpful to make image accessibility a quality dimension within Wikipedia. We estimate that more than 40% of the English Wikipedia articles marked by the community as featured that is of best quality have no alt texts for any of their images. Further providing accessibility descriptions could become more central in the editing tools, for example by requiring contributors to write alt text for every new image before saving the content. Finally, direct community engagement can focus on content that needs urgent attention and connect contributors with blind and low vision people who can offer valuable feedback on their needs. The additional implementation of tools to surface articles and images without accessibility coverage and support editors and the difficult task of writing alt text could help improve quality. Lastly mixed human AI systems may prove effective by lowering the participation barrier, since it is often easier to edit an existing text than it is to write one from scratch. We also may be able to find areas of Wikipedia that are unusually successful in achieving good coverage for accessibility descriptions and such communities could teach us important lessons. In summary, our work investigates accessibility of images on Wikipedia and outline steps towards addressing this challenge. Thank you.