Home

Awesome

K-BERT

Sorce code and datasets for "K-BERT: Enabling Language Representation with Knowledge Graph", which is implemented based on the UER framework.

News

Requirements

Software:

Python3
Pytorch >= 1.0
argparse == 1.1

Prepare

The directory tree of K-BERT:

K-BERT
├── brain
│   ├── config.py
│   ├── __init__.py
│   ├── kgs
│   │   ├── CnDbpedia.spo
│   │   ├── HowNet.spo
│   │   └── Medical.spo
│   └── knowgraph.py
├── datasets
│   ├── book_review
│   │   ├── dev.tsv
│   │   ├── test.tsv
│   │   └── train.tsv
│   ├── chnsenticorp
│   │   ├── dev.tsv
│   │   ├── test.tsv
│   │   └── train.tsv
│    ...
│
├── models
│   ├── google_config.json
│   ├── google_model.bin
│   └── google_vocab.txt
├── outputs
├── uer
├── README.md
├── requirements.txt
├── run_kbert_cls.py
└── run_kbert_ner.py

K-BERT for text classification

Classification example

Run example on Book review with CnDbpedia:

CUDA_VISIBLE_DEVICES='0' nohup python3 -u run_kbert_cls.py \
    --pretrained_model_path ./models/google_model.bin \
    --config_path ./models/google_config.json \
    --vocab_path ./models/google_vocab.txt \
    --train_path ./datasets/book_review/train.tsv \
    --dev_path ./datasets/book_review/dev.tsv \
    --test_path ./datasets/book_review/test.tsv \
    --epochs_num 5 --batch_size 32 --kg_name CnDbpedia \
    --output_model_path ./outputs/kbert_bookreview_CnDbpedia.bin \
    > ./outputs/kbert_bookreview_CnDbpedia.log &

Results:

Best accuracy in dev : 88.80%
Best accuracy in test: 87.69%

Options of run_kbert_cls.py:

useage: [--pretrained_model_path] - Path to the pre-trained model parameters.
        [--config_path] - Path to the model configuration file.
        [--vocab_path] - Path to the vocabulary file.
        --train_path - Path to the training dataset.
        --dev_path - Path to the validating dataset.
        --test_path - Path to the testing dataset.
        [--epochs_num] - The number of training epoches.
        [--batch_size] - Batch size of the training process.
        [--kg_name] - The name of knowledge graph, "HowNet", "CnDbpedia" or "Medical".
        [--output_model_path] - Path to the output model.

Classification benchmarks

Accuracy (dev/test %) on different dataset:

DatasetHowNetCnDbpedia
Book review88.75/87.7588.80/87.69
ChnSentiCorp95.00/95.5094.42/95.25
Shopping97.01/96.9296.94/96.73
Weibo98.22/98.3398.29/98.33
LCQMC88.97/87.1488.91/87.20
XNLI77.11/77.0776.99/77.43

K-BERT for named entity recognization (NER)

NER example

Run an example on the msra_ner dataset with CnDbpedia:

CUDA_VISIBLE_DEVICES='0' nohup python3 -u run_kbert_ner.py \
    --pretrained_model_path ./models/google_model.bin \
    --config_path ./models/google_config.json \
    --vocab_path ./models/google_vocab.txt \
    --train_path ./datasets/msra_ner/train.tsv \
    --dev_path ./datasets/msra_ner/dev.tsv \
    --test_path ./datasets/msra_ner/test.tsv \
    --epochs_num 5 --batch_size 16 --kg_name CnDbpedia \
    --output_model_path ./outputs/kbert_msraner_CnDbpedia.bin \
    > ./outputs/kbert_msraner_CnDbpedia.log &

Results:

The best in dev : precision=0.957, recall=0.962, f1=0.960
The best in test: precision=0.953, recall=0.959, f1=0.956

Options of run_kbert_ner.py:

useage: [--pretrained_model_path] - Path to the pre-trained model parameters.
        [--config_path] - Path to the model configuration file.
        [--vocab_path] - Path to the vocabulary file.
        --train_path - Path to the training dataset.
        --dev_path - Path to the validating dataset.
        --test_path - Path to the testing dataset.
        [--epochs_num] - The number of training epoches.
        [--batch_size] - Batch size of the training process.
        [--kg_name] - The name of knowledge graph.
        [--output_model_path] - Path to the output model.

K-BERT for domain-specific tasks

Experimental results on domain-specific tasks (Precision/Recall/F1 %):

KGFinance_QALaw_QAFinance_NERMedicine_NER
HowNet0.805/0.888/0.8450.842/0.903/0.8710.860/0.888/0.8740.935/0.939/0.937
CN-DBpedia0.814/0.881/0.8460.814/0.942/0.8740.860/0.887/0.8730.935/0.937/0.936
MedicalKG------0.944/0.943/0.944

Acknowledgement

This work is a joint study with the support of Peking University and Tencent Inc.

If you use this code, please cite this paper:

@inproceedings{weijie2019kbert,
  title={{K-BERT}: Enabling Language Representation with Knowledge Graph},
  author={Weijie Liu, Peng Zhou, Zhe Zhao, Zhiruo Wang, Qi Ju, Haotang Deng, Ping Wang},
  booktitle={Proceedings of AAAI 2020},
  year={2020}
}