This is the official implementation for MoCA, which focuses on the task of Textbook Question Answering.

MoCA: Incorporating Multi-stage Domain Pretraining and Cross-guided Multimodal Attention for Textbook Question Answering
(Pattern Recognition) [paper]

For the external corpus and some important checkpoints, please download them from here.

• Multi-stage Pretrain for text part
• Dense Layer of Text-guided Visual Attention for diagram part

Conducted on single Tesla-v100

If you find it helpful, please kindly cite the paper.

@article{xu2021moca,
  title   =   {MoCA: Incorporating Multi-stage Domain Pretraining and Cross-guided Multimodal Attention for Textbook Question Answering},
  author  =   {Xu, Fangzhi and Lin, Qika and Liu, Jun and Zhang, Lingling and Zhao, Tianzhe and Chai, Qi and Pan, Yudai},
  journal =   {arXiv preprint arXiv:2112.02839},
  year    =   {2021}
}