chunked/nltk_chunking/_model_summary/chunk_85.txt

VisualBERT predicts the masked text based on the unmasked text and the visual embeddings, and it also has to predict whether the text is aligned with the image.