Pivot-based entity linking (PBEL) is an entity linking method that uses high-resource languages as intermediates while doing low-resource cross-lingual entity linking. The details are described in the paper: Zero-shot Neural Transfer for Cross-lingual Entity Linking.
- Python2/3 with NumPy
- Dynet
- PanPhon
- Epitran for converting strings to IPA (needed if creating your own data)
- Download sample data for one language pair here.
- Download data for all 54 training languages and 9 test languages here.
train.pyis used for training the entity similarity model.- Default values for hyperparameters are in the script -- 64 size character embeddings and 1024 hidden size for the character LSTM.
test.pyis used both for encoding the knowledge base as well as retrieving entity linking candidates for test data or other input files.- See the
scripts/folder for examples using data from here.
If you use this repository, please cite
@inproceedings{rijhwani19aaai,
title = {Zero-shot Neural Transfer for Cross-lingual Entity Linking},
author = {Shruti Rijhwani and Jiateng Xie and Graham Neubig and Jaime Carbonell},
booktitle = {Thirty-Third AAAI Conference on Artificial Intelligence (AAAI)},
address = {Honolulu, Hawaii},
month = {January},
url = {https://arxiv.org/abs/1811.04154},
year = {2019}
}