Wikidata-Descriptions

Be Concise and Precise: Synthesizing Open-Domain Entity Descriptions from Facts

Training and Inference

python run_model.py --model MODEL_NAME --datadir PATH_TO_DATADIR --config PATH_TO_CONFIG_FILE --mode [training/eval]

MODEL_NAME can be NKLM, DGN, Fact2SeqAttn, or PointerGenerator. E.g.

run_model.py --model NKLM --datadir Data/WikiFacts10k-OpenDomain --config Configs/NKLM-config.json --mode training
run_model.py --model PointerGenerator --datadir Data/WikiFacts10k-OpenDomain --config Configs/PointerGenerator-config.json --mode eval

NOTE: NKLM requires pre-trained TransE embeddings of entities and relations. To train the TransE model, run the following:

python TransE.py --datadir PATH_TO_DATADIR

Evaluation

We used the automatic evaluation script developed by Sharma et al., 2017. To run the evaluation script, run the following:

python evaluation.py <model_name>.csv

Authors

• Rajarshi Bhowmik - website
• Gerard de Melo - website

Citation

If you use this code, please cite our paper.

@inproceedings{BhowmikDeMelo2019EntityDescriptionsCopyModel,
  title = {Be Concise and Precise: Synthesizing Open-Domain Entity Descriptions from Facts},
  author = {Bhowmik, Rajarshi and {de Melo}, Gerard},
  booktitle = {Proceedings of The Web Conference 2019},
  year = {2019},
  location = {San Francisco},
}

License

This project is licensed under the MIT License - see the LICENSE file for details