The SentiCap dataset contains images with positive and negative sentiment descriptions. These sentiment descriptions were generated by the authors by rewriting factual descriptions. There are more than 2,000 sentiment descriptions in total.

The images in the SentiCap dataset are mainly taken from the MS COCO dataset. Images are annotated based on the polarity of emotions, and "positive" and "negative" image descriptions are formed for each image. This dataset uses 1,027 positive ANPs (adjective noun pairs) and 436 negative ANPs. The dataset ultimately contains a total of 3,171 images, and each image corresponds to at least 3 positive and 3 negative statement descriptions.