CrossWeigh

CrossWeigh

CrossWeigh: Training Named Entity Tagger from Imperfect Annotations

Motivation

The label annotation mistakes by human annotators brings up two challenges to NER:

• mistakes in the test set can interfere the evaluation results and even lead to an inaccurate assessment of model performance.
• mistakes in the training set can hurt NER model training.

We address these two problems by:

• manually correcting the mistakes in the test set to form a cleaner benchmark.
• develop framework CrossWeigh to handle the mistakes in the training set.

CrossWeigh works with any NER algorithm that accepts weighted training instances. It is composed of two modules. 1) mistake estimation: where potential mistakes are identified in the training data through a cross-checking process and 2) mistake re-weighing: where weights of those mistakes are lowered during training the final NER model.

Data

We formally name our corrected dataset as CoNLL++.
/data/conllpp_test.txt is the manually corrected test set, there should be exactly 186 sentences that differ from the original test set.
/data/conllpp_train.txt and /data/conllpp_dev.txt are the original dataset of CoNLL03 from Named-Entity-Recognition-NER-Papers.

Scripts

split.py can be used to generate a k-fold entity disjoint dataset from a list of datasets(usually both the train and development set)
flair_scripts/flair_ner.py can be used to train a weighted version of flair.
collect.py can be used to collect all the predictions on the k folded test set.

Steps to reproduce

Make sure you are in a python3.6+ environment.
See example.sh to reproduce the results.
Using Flair (non-pooled version), the final result should achieve around 93.19F1 on the original test dataset and 94.18F1 on the corrected test set. Using Flair without CrossWeigh gives around 92.9F1.

Results

All the results are averaged across 5 runs and standard deviation is reported.

For all models, we use their suggested parameter settings.
For GCDT, we used the weights estimated from Pooled-Flair for efficiency purposes.

Citation

Please cite the following paper if you found our dataset or framework useful. Thanks!

Zihan Wang, Jingbo Shang, Liyuan Liu, Lihao Lu, Jiacheng Liu, and Jiawei Han. "CrossWeigh: Training Named Entity Tagger from Imperfect Annotations." arXiv preprint arXiv:1909.01441 (2019).

@article{wang2019cross,
  title={CrossWeigh: Training Named Entity Tagger from Imperfect Annotations},
  author={Wang, Zihan and Shang, Jingbo and Liu, Liyuan and Lu, Lihao and Liu, Jiacheng and Han, Jiawei},
  journal={arXiv preprint arXiv:1909.01441},
  year={2019}
}