Large-scale visual-linguistic pre-training aims to capture the generic representations from multimodal features, which are essential for downstream vision-language tasks. Existing methods mostly focus on learning the semantic connections between visual objects and linguistic content, which tend to be recognitionlevel information and may not be sufficient for commonsensical reasoning tasks like VCR. In this paper, we propose a novel commonsensical vision-language pre-training framework to bridge the gap. We first augment the conventional image-caption pre-training datasets with commonsense inferences from a visuallinguistic GPT-2. To pre-train models on image, caption and commonsense inferences together, we propose two new tasks: masked commonsense modeling (MCM) and commonsense type prediction (CTP). To reduce the shortcut effect between captions and commonsense inferences, we further introduce the domain-wise adaptive masking that dynamically adjusts the masking ratio. Experimental results on downstream tasks, VCR and VQA, show the improvement of our pre-training strategy over previous methods. Human evaluation also validates the relevance, informativeness, and diversity of the generated commonsense inferences. Overall, we demonstrate the potential of incorporating commonsense knowledge into the conventional recognition-level visual-linguistic pre-training.

Pre-trained contextual vision-and-language (V&L) models have achieved impressive performance on various benchmarks. However, existing models require a large amount of parallel image-caption data for pre-training. Such data are costly to collect and require cumbersome curation. Inspired by unsupervised machine translation, we investigate if a strong V&L representation model can be learned through unsupervised pre-training without image-caption corpora. In particular, we propose to conduct “mask-and-predict” pre-training on text-only and image-only corpora and introduce the object tags detected by an object recognition model as anchor points to bridge two modalities. We find that such a simple approach achieves performance close to a model pre-trained with aligned data, on four English V&L benchmarks. Our work challenges the widely held notion that aligned data is necessary for V&L pre-training, while significantly reducing the amount of supervision needed for V&L models.