BOND

This repo contains our code and pre-processed distantly/weakly labeled data for paper BOND: BERT-Assisted Open-Domain Name Entity Recognition with Distant Supervision (KDD2020)

BOND

Benchmark

The reuslts (entity-level F1 score) are summarized as follows:

| Method | CoNLL03 | Tweet | OntoNote5.0 | Webpage | Wikigold | | ------ | ------- | ----- | ----------- | ------- | -------- | | Full Supervision | 91.21 | 52.19 | 86.20 | 72.39 | 86.43 | | Previous SOTA | 76.00 | 26.10 | 67.69 | 51.39 | 47.54 | | BOND | 81.48 | 48.01 | 68.35 | 65.74 | 60.07 |

• Full Supervision: Roberta Finetuning/BiLSTM CRF
• Previous SOTA: BiLSTM-CRF/AutoNER/LR-CRF/KALM/CONNET

Data

We release five open-domain distantly/weakly labeled NER datasets here: dataset. For gazetteers information and distant label generation code, please directly email cliang73@gatech.edu.

Environment

Python 3.7, Pytorch 1.3, Hugging Face Transformers v2.3.0.

Training & Evaluation

We provides the training scripts for all five open-domain distantly/weakly labeled NER datasets in scripts. E.g., for BOND training and evaluation on CoNLL03

cd BOND
./scripts/conll_self_training.sh

For Stage I training and evaluation on CoNLL03

cd BOND
./scripts/conll_baseline.sh

The test reuslts (entity-level F1 score) are summarized as follows:

| Method | CoNLL03 | Tweet | OntoNote5.0 | Webpage | Wikigold | | ------ | ------- | ----- | ----------- | ------- | -------- | | Stage I| 75.61 | 46.61 | 68.11 | 59.11 | 52.15 | | BOND | 81.48 | 48.01 | 68.35 | 65.74 | 60.07 |

Citation

Please cite the following paper if you are using our datasets/tool. Thanks!

@inproceedings{liang2020bond,
  title={BOND: Bert-Assisted Open-Domain Named Entity Recognition with Distant Supervision},
  author={Liang, Chen and Yu, Yue and Jiang, Haoming and Er, Siawpeng and Wang, Ruijia and Zhao, Tuo and Zhang, Chao},
  booktitle={ACM SIGKDD International Conference on Knowledge Discovery and Data Mining},
  year={2020}
}