Skip to content

Gengchen Mai, Krzysztof Janowicz, Ling Cai, Rui Zhu, Blake Regalia, Bo Yan, Meilin Shi, Ni Lao. SE-KGE: A Location-Aware Knowledge Graph Embedding Model for Geographic Question Answering and Spatial Semantic Lifting. Transactions in GIS. DOI:10.1111/TGIS.12629

master
Switch branches/tags
Code

Latest commit

 

Git stats

Files

Permalink
Failed to load latest commit information.
Type
Name
Latest commit message
Commit time
 
 
img
 
 
 
 
 
 
 
 
 
 

Geographic Question Answering based on Knowledge Graph Embedding Techniques

Code for recreating the results in our Transactions in GIS paper as well as our K-CAP 2019 paper .

SE‐KGE : A location‐Aware Knowledge Graph Embedding Model for Geographic Question Answering and Spatial Semantic Lifting

Related Link

  1. Transactions in GIS Paper
  2. Arxiv Paper
  3. Esri UC 2020 Presentation
  4. Esri UC 2020 Presentation Slides

Contextual Graph Attention for Answering Logical Queries over Incomplete Knowledge Graphs

Related Link

  1. K-CAP 2019 Paper
  2. Arxiv Paper
  3. K-CAP 2019 Presentation Slides

Please visit my Homepage for more information.

Model Archtecture

Geographic Question Answering

geoqa

The location-aware entity encoder architecture:

ent_enc

Spatial Semantic Lifting

spa_sem_lift

Dependencies

  • Python 2.7+
  • Torch 1.0.1+
  • numpy 1.16.0+
  • matplotlib 2.2.4+
  • sklearn 0.20.3+
  • geopandas 0.6.1+
  • shapely 1.6.4+
  • pyproj 2.2.2+

To set up the code , run python setup.py. Note that the first three are required for model training and testing. The rest are used for visualization which is optional.

Data

You can download the GeoQA dataset from here. Unextract it and put them in graphqa/dbgeo/.

Code Usage

This code is implemented in Python 2.7. All codes are in graphqa/netquery/.

Geographic Question Answering

For each baseline in Table 3:

  1. GQE_{diag}: run graphqa/dbgeo_geoqa_gqe.sh;
  2. GQE: run graphqa/dbgeo_geoqa_gqe_diag.sh;
  3. CGA: run graphqa/dbgeo_geoqa_cga.sh;
  4. SE-KGE_{direct}: run graphqa/dbgeo_geoqa_direct.sh;
  5. SE-KGE_{pt}: run graphqa/dbgeo_geoqa_direct.sh;
  6. SE-KGE_{space}: run graphqa/dbgeo_geoqa_space.sh;
  7. SE-KGE_{full}: run graphqa/dbgeo_geoqa_full.sh.

Spatial Semantic Lifting

For each baseline in Table 5:

  1. SE-KGE_{space}: run graphqa/dbgeo_spa_sem_lift_space.sh;
  2. SE-KGE_{ssl}: run graphqa/dbgeo_spa_sem_lift_ssl.sh.

Reference

If you find our work useful in your research please consider citing our paper.

@article{mai2020se,
  title={{SE}-{KGE}: A Location-Aware Knowledge Graph Embedding Model for Geographic Question Answering and Spatial Semantic Lifting},
  author={Mai, Gengchen and Janowicz, Krzysztof and Cai, Ling and Zhu, Rui and Regalia, Blake and Yan, Bo and Shi, Meilin and Lao, Ni},
  journal={Transactions in GIS},
  year={2020},
  doi={10.1111/tgis.12629}
}

@inproceedings{mai2019contextual,
  title={Contextual Graph Attention for Answering Logical Queries over Incomplete Knowledge Graphs},
  author={Mai, Gengchen and Janowicz, Krzysztof and Yan, Bo and Zhu, Rui and Cai, Ling and Lao, Ni},
  booktitle={Proceedings of the 10th International Conference on Knowledge Capture},
  pages={171--178},
  year={2019}
}

The location encoder component in SE-KGE model is based on Space2Vec. Read our ICLR 2020 paper for a comprehensive understanding:

@inproceedings{space2vec_iclr2020,
  title={Multi-Scale Representation Learning for Spatial Feature Distributions using Grid Cells},
  author={Mai, Gengchen and Janowicz, Krzysztof and Yan, Bo and Zhu, Rui and  Cai, Ling and Lao, Ni},
  booktitle={The Eighth International Conference on Learning Representations},
  year={2020},
  organization={openreview}
}

Note that a part of our code is based on the code of Hamilton et al's NIPS 2018 paper:

@inproceedings{hamilton2018embedding,
  title={Embedding logical queries on knowledge graphs},
  author={Hamilton, Will and Bajaj, Payal and Zitnik, Marinka and Jurafsky, Dan and Leskovec, Jure},
  booktitle={Advances in Neural Information Processing Systems},
  pages={2026--2037},
  year={2018}
}

About

Gengchen Mai, Krzysztof Janowicz, Ling Cai, Rui Zhu, Blake Regalia, Bo Yan, Meilin Shi, Ni Lao. SE-KGE: A Location-Aware Knowledge Graph Embedding Model for Geographic Question Answering and Spatial Semantic Lifting. Transactions in GIS. DOI:10.1111/TGIS.12629

Resources

License

Releases

No releases published

Packages

No packages published