TRADE: TRuly ADversarial ad understanding Evaluation

TRADE (TRuly ADversarial ad understanding Evaluation) is a new diagnostic test set consisting of 300 randomly selected image-based advertisements from the Pitt Ads dataset. Each ad is associated with 3 options: one valid explanation from the available annotations in the Pitt Ads dataset and two newly created adversarial negative explanations. TRADE is presented in the following publication; please cite it if you use the dataset in your work: Anna Bavaresco, Alberto Testoni, and Raquel Fernández, 2024. Don't Buy it! Reassessing the Ad Understanding Abilities of Contrastive Multimodal Models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL-2024). Association for Computational Linguistics. The paper can be found here: https://aclanthology.org/2024.acl-short.77/