Data-driven accessibility repair revisited: on the effectiveness of generating labels for icons in Android apps

Mobile apps are playing an increasingly important role in our daily lives, including the lives of approximately 304 million users worldwide that are either completely blind or suffer from some form of visual impairment. These users rely on screen readers to interact with apps. Screen readers, however, cannot describe the image icons that appear on the screen, unless those icons are accompanied with developer-provided textual labels. A prior study of over 5,000 Android apps found that in around 50% of the apps, less than 10% of the icons are labeled. To address this problem, a recent award-winning approach, called LabelDroid, employed deep-learning techniques to train a model on a dataset of existing icons with labels to automatically generate labels for visually similar, unlabeled icons. In this work, we empirically study the nature of icon labels in terms of distribution and their dependency on different sources of information. We then assess the effectiveness of LabelDroid in predicting labels for unlabeled icons. We find that icon images are insufficient in representing icon labels, while other sources of information from the icon usage context can enrich images in determining proper tokens for labels. We propose the first context-aware label generation approach, called COALA, that incorporates several sources of information from the icon in generating accurate labels. Our experiments show that although COALA significantly outperforms LabelDroid in both user study and automatic evaluation, further research is needed. We suggest that future studies should be more cautious when basing their approach on automatically extracted labeled data.