Improving Candidate Generation for Low-resource Cross-lingual Entity Linking

This is the code repo for TACL 2020 paper Improving Candidate Generation for Low-resource Cross-lingual Entity Linking

Training

Please refer to train.sh for the arguments, all four models (charagram, charcnn, lstm-last and lstm-avg) could be launched in this bash file

Test

Please refer to test.sh for the arguments, all four models (charagram, charcnn, lstm-last and lstm-avg) could be launched in this bash file

Data

Data folder contains data alias and kb

data: data for train, dev and test

English_Wikipedia_ID ||| English_Wikipedia_title ||| Wikipedia_title_of_train/test_lang ||| Entity_type
e.g. 3378263 ||| John Michael Talbot ||| ጆን ማይክል ታልበት ||| PER

alias: entity alias from Wikidata

Wikidata_ID ||| English_Wikipedia_ID ||| aliases
e.g. Q42 ||| 8091 ||| Douglas Adams ||| Douglas N. Adams || Douglas Noël Adams || Douglas Noel Adams

kb: all entities in wikipedia (~2M, proper nouns)

English_Wikipedia_ID ||| English_Wikipedia_title ||| Entity_type
e.g. 16429160 ||| George Karrys ||| PER

Note that all files ended with .ipa are of phoneme representations, using Epitran

Model

under the models folder, we have:

base_train: the base class for training

base_test: the base class for testing

charagram: character n-gram model, used in our work

charcnn: character CNN, used as baseline

lstm: used as baseline

Utils

to_ipa.py: code for generating the phoneme representation of strings.

TODO

• Trained models
• Code to extract data from Wikipedia for any language and specify train/dev/test split
• Code to calculate n-gram

Reference

@article{zhou20tacl,
    title = {Improving Candidate Generation for Low-resource Cross-lingual Entity Linking},
    author = {Shuyan Zhou and Shruti Rijhwani and John Wieting and Jaime Carbonell and Graham Neubig},
    journal = {Transactions of the Association of Computational Linguistics},
    month = {},
    url = {https://arxiv.org/abs/2003.01343},
    year = {2020}
}