Home

Awesome

MGeo

MGeo: Multi-Modal Geographic Language Model Pre-Training

Release Note

Download

Reproduce results in paper

Prepare environment

conda create -n mgeo python=3.7
pip install -r requirements.txt

Download resources

cd data
unzip datasets.zip
cd ../prepare_data
download_pretrain_models.sh

Generate pretrain data

We only provide samples of pretrain data. To produce your own pretrain data, you simply need text-geolocation pairs which can be genrate by various way (e.g., user click, POI data, position of delivery clerks). The geolocation and text just need to be related, no need to be exactly precise.

Having text-geolocation pairs, you can follow steps below to generate pretrain data. Demo pairs are saved in resources/text_location_pair.demo for testing.

Pretrain geographic encoder

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 sh run_gis_encoder_pretrain.sh

Pretrain multimodal interaction module

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 sh run_mm_pretrain.sh

Finetune on rerank task

CUDA_VISIBLE_DEVICES=0,1,2,3 sh run_rerank.sh

Finetune on retrieval task

CUDA_VISIBLE_DEVICES=0,1,2,3 sh run_retrieval.sh

Contact

Please contact ada.drx@alibaba-inc.com to get pretrained model or other resources.

Reference

@article{ding2023multimodal,
  title={MGeo: A Multi-Modal Geographic Pre-Training Method},
  author={Ruixue Ding and Boli Chen and Pengjun Xie and Fei Huang and Xin Li and Qiang Zhang and Yao Xu},
  journal={Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval},
  year={2023}
}