MGeo

MGeo: Multi-Modal Geographic Language Model Pre-Training

Release Note

2023.01.16 Pretrained model and datasets are released in modelscope
2023.04.05 Paper is accepted by SIGIR2023. Paper link: https://arxiv.org/abs/2301.04283
2023.05.17 Pretraining and finetune codes in paper are released
2023.07.06 Pretrained model used to reproduce paper is released
2023.07.25 dataset.zip download url is provided
2023.09.14 Finetuned model download is provided

Download

Pretrained model and finetune code for a more general geographic model on 6 tasks are availabel at https://modelscope.cn/models/damo/mgeo_backbone_chinese_base/files , notice this model is different from the paper.
Datasets for 6 tasks are available at https://modelscope.cn/datasets/damo/GeoGLUE/files, notice this dataset is different from the paper.
Pretrained model used to reproduce paper results: https://drive.google.com/file/d/1j6S52jkxks4UBsCU8ZgLroqksUYSz5x7/view?usp=sharing
dataset.zip in paper is available at https://drive.google.com/file/d/1KxN377_l-yNwsYBifS_o0f50tKBYNYj-/view?usp=sharing
Finetuned models in paper are available at https://drive.google.com/file/d/1pIn8aemh2VyICG32VbWVthzXpsLSSSvp/view?usp=sharing

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.

Download proper map from OpenStreetMap. For example, our geolocations used in paper are in HangZhou. Thus, we download China map.
Import map and your text-geolocation data to a GIS database, like PostGIS. Every text is assigned with an ID. In demo case, ID is row number.
Find COVERED and NEARBY relations between map and geolocation using GIS database. Export to resources/location_in_aoi.demo, resources/location_near_aoi.demo, resources/location_near_road.demo.
Export needed AOIs and roads to resources/hz_aoi.txt and resources/hz_roads.txt. In our paper's setting, only elements in HangZhou are included.
Generate pretrain data using command: cd prepare_data && python calculate_geographic_context.py . Use the length of geom_ids in calculate_geographic_context.py to replace vocab_size in resources/gis_config.json.

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}
}

PhantomGrapes/MGeo