• We conduct analyses of the performance of existing VQA model with respect to logically composed questions


• We curate two large scale datasets VQA-Compose and VQA-Supplement that contain logically composed binary questions.


• We show a capability of answering logically composed questions with our novel modules, while retaining performance on VQA data.

VQA models struggle at negation, antonyms, conjunction, disjunction!

We provide dataset, a new model, and detailed analysis!

Our dataset contains questions composed with negation, antonyms, conjunctions, and disjunctions.
Our model learns to identify the type of question and they type of logical connective in the question to aid question-answering.

Reference

Gokhale, T., Banerjee, P., Baral, C., & Yang, Y. (2020). VQA-LOL: Visual Question Answering under the Lens of Logic.
Bibtex