Czerwionka, Thomas, and Claudia De Witt. Betreuung von Online-Communities of inquiry. na, 2006.

This paper introduces VQAonline, the first Visual Question Answering (VQA) dataset derived from real-world scenarios, specifically online community forums. Unlike traditional VQA datasets, VQAonline features answers that are significantly longer, averaging 173 words. Given this unique characteristic, the paper evaluates six popular metrics for longer text evaluation to determine their alignment with human judgments. After identifying the most suitable metrics, the study assesses six leading vision and language foundation models using VQAonline, pinpointing their main challenges. The dataset is publicly available, offering a valuable resource for advancing research in the VQA field by providing a more authentic and complex set of data for training and evaluating AI models.