This is the dataset proposed in the EMNLP 2024 paper: Encoding and Controlling Global Semantics for Long-form Video Question Answering.
The dataset consists of 19k video question answering data whose domain is movies. We create the dataset via 2 stages:
Question-answer generation: for each video, we concatenate the video captions following the time order to construct the video description. Then, we utilize GPT-4 to generate 20 questions per video.
Data filtering: we… See the full description on the dataset page:
https://huggingface.co/datasets/thongnguyen5999/madqa.
