Hybrid Multi-Aspect Alignment Networks (HMAN)

The code for the research paper Aligning Cross-lingual Entities with Multi-Aspect Information @EMNLP 2019.

*Our code is built on top of GCN-Align.

Quick Demo

We stored the embeddings for demo in the directory demo_embd, and we can evaluate ZH-EN as follows:

python weighted_concat.py -d demo_embd/pairwise_dump.json -g demo_embd/zh_en_graph_embd.pkl -i data/zh_en/test

Run Graph-based Embeddings (HMAN/MAN)

HMAN

bash graph.sh 0 1

MAN

bash graph.sh 0 0

Run PairwiseBERT

We use the package relogic to derive BERT-based embeddings. Please install relogic first:

git submodule init
git submodule update

Also, becasuerelogic evolves fast, we suggest that you change to the commit d1b5046:

cd relogic
git checkout d1b5046

Note that the argument --local_rank in relogic indicates your gpu id.

Training

bash train_bert.sh 0 zh_en

(Note that you need to stop the training manually.)

Evaluation

bash eval_bert.sh 0 zh_en

Integration

python weighted_concat.py --desc relogic/saves/pair_matching/zh_en/pairwise_dump.json --graph graph_ckpt/zh_en_graph_embd.pkl --ill data/zh_en/test

Citation

@article{yang2019aligning,
  title={Aligning Cross-Lingual Entities with Multi-Aspect Information},
  author={Yang, Hsiu-Wei and Zou, Yanyan and Shi, Peng and Lu, Wei and Lin, Jimmy and Sun, Xu},
  booktitle={Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing},
  year={2019}
}

h324yang/HMAN