Leverage Lexical Knowledge for Chinese Named Entity Recognition via Collaborative Graph Network
Source code for Leverage Lexical Knowledge for Chinese Named Entity Recognition via Collaborative Graph Network in EMNLP 2019. If you use this code or our results in your research, we would appreciate it if you cite our paper as following:
@article{Sui2019Graph4CNER,
title = {Leverage Lexical Knowledge for Chinese Named Entity Recognition via Collaborative Graph Network},
author = {Sui, Dianbo and Chen, Yubo and Liu, Kang and Zhao, Jun and Liu, Shengping},
journal = {Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing},
year = {2019}
}
Requirements:
Python: 3.7
PyTorch: 1.1.0
Input format:
Input is in CoNLL format (We use BIO tag scheme), where each character and its label are in one line. Sentences are split with a null line.
叶 B-PER
嘉 I-PER
莹 I-PER
先 O
生 O
获 O
聘 O
南 B-ORG
开 I-ORG
大 I-ORG
学 I-ORG
终 O
身 O
校 O
董 O
。 O
Pretrained Embeddings:
Character embeddings (gigaword_chn.all.a2b.uni.ite50.vec) can be downloaded in Google Drive or Baidu Pan.
Word embeddings (sgns.merge.word) can be downloaded in Google Drive or Baidu Pan.
Usage：
1️⃣ Download the character embeddings and word embeddings and put them in the data/embeddings folder.
2️⃣ Modify the run_main.sh by adding your train/dev/test file directory.
3️⃣ sh run_main.sh. Note that the default hyperparameters is may not be the optimal hyperparameters, and you need to adjust these.
:four: Enjoy it! :smile:
Result：
For WeiboNER dataset，using the default hyperparameters in run_main.sh can achieve the state-of-art results (Dev F1: 66.12%; Test F1: 66.84%). 😎
Speed：
I have optimized the code and this version is faster than the one in our paper. 💪