It implements:
| Item | Source/Reference | |
| Models | ||
| BiLSTM-Linear | Long Short-term Memory | |
| BiLSTM-Linear-CRF | Neural Architectures for Named Entity Recognition | |
| BERT-Linear | BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding | |
| BERT-Linear-CRF | ||
| BERT-BiLSTM-Linear | ||
| BERT-BiLSTM-Linear-CRF | ||
| BERT(Prompt) EntLM Approach | Template-free Prompt Tuning for Few-shot NER | |
| Datasets | ||
| CoNLL2003 | yuanxiaosc/BERT-for-Sequence-Labeling-and-Text-Classification | |
| OntoNotes5 | LDC2013T19 | |
| CCKS2019 Subtask 1 | TIANCHI (NER on Chinese medical documents.) | |
| NCBI-disease | BioBERT (NER on English medical doucments, Got these datasets from its download.sh) | |
| Traning Tricks | ||
| Gradient Accumulation | ||
| Learning Rate Warmup | ||
| Misc | ||
| Tokenizer from datasets | See myutils.py | |
| NER Metrics | seqeval: A Python framework for sequence labeling evaluation |
You can easily add your own models and datasets into this framework.
Linux(Tested)/Windows(Not Tested) with Nvidia GPUs.
Recommend to use conda creating a python environment(python==3.9). For example:
conda create -n NER python=3.9And run the bash script. If you are using windows, change its extname to .bat.
./install_dependencies.sh
For some reason(copyright and some other things), I can't directly provide datasets to you. You should get the access to these datasets by yourself and put them in specified format into 'assert/raw_datasets' folder, see here.
| Optimizer | Weight Decay |
Warmup Ratio |
Batch Size | Gradient Accumulation |
Clip Grad Norm | Random Seed |
| AdamW | 5e-3 | 0.2 | 1 | 32 | 1.0 | 233 |
Training Epoches:
| Dataset | Full Data | Few Shot |
| CoNLL2003 | 12 | 30 |
| OntoNotes5(Chinese) | ||
| CCKS2019 | ||
| NCBI-disease | 20 |
Learning Rates:
| CoNLL2003 | OntoNotes5 | CCKS2019 | NCBI-disease | |
| BiLSTM-Linear | 0.001 | NA | ||
| BiLSTM-Linear-CRF | ||||
| BERT-Linear | 0.0001 | |||
| BERT-Linear-CRF | ||||
| BERT-BiLSTM-Linear | 0.0001 | 3e-5 | NA | |
| BERT-BiLSTM-Linear-CRF | 1e-5 | |||
| BERT(Prompt) | 3e-5 | 0.0001 | ||
| BERT Model | Embedding Size(For models without BERT) | LSTM Hidden Size | LSTM Layers |
| bert-base-uncased(CoNLL2003,NCBI-disease) | 256 | 256 | 2 |
| bert-base-chinese(OntoNotes5,CCKS2019) |
General datasets(ConLL2003, OntoNotes5(Chinese)).
| Dataset | Model | Overall Span-Based Micro F1 | Average Training Time Per Epoch (On a Quadro RTX8000) |
| CoNLL2003 | BiLSTM-Linear | 0.6517005491858561 | 13.98s |
| BiLSTM-Linear-CRF | 0.6949365863103882 | 44.07s | |
| BERT-Linear | 0.8983771483322356 | 81.81s | |
| BERT-Linear-CRF | 0.8977943835121128 | 120.94s | |
| BERT-BiLSTM-Linear | 0.8819152766110644 | 117.37s | |
| BERT-BiLSTM-Linear-CRF | 0.8873846891098599 | 130.85s | |
| BERT(Prompt) | 0.9230769230769231 | 99.70s | |
| OntoNotes5 (Chinese) | BiLSTM-Linear | 0.637999350438454 | 160.55s |
| BiLSTM-Linear-CRF | 0.7033358449208851 | 319.87s | |
| BERT-Linear | 0.7403041825095057 | 413.20s | |
| BERT-Linear-CRF | 0.7535838822161953 | 595.71s | |
| BERT-BiLSTM-Linear | 0.7511438739196745 | 590.53s | |
| BERT-BiLSTM-Linear-CRF | 0.7616389699353039 | 800.23s | |
| BERT(Prompt) | 0.7376454875023851 | 485.56s |
Medical datasets, used general bert and medical bert.
| Dataset | BERT | Model | Overall Span-Based Micro F1 |
| CCKS2019 Subtask1 |
bert- base- chinese |
BERT-Linear | 0.8057400574005741 |
| BERT-Linear-CRF | 0.8119778310861113 | ||
| BERT-Prompt | 0.7684884784959654 | ||
| medbert- base- chinese |
BERT-Linear | 0.8201214508452324 | |
| BERT-Linear-CRF | 0.8221622063998691 | ||
| BERT-Prompt | 0.7933091394485463 | ||
| NCBI-disease | bert- base- uncased |
BERT-Linear | 0.8720903433970961 |
| BERT-Linear-CRF | 0.8778661675245673 | ||
| BERT-Prompt | 0.8319672131147542 | ||
|
biobert-base- cased-v1.2 |
BERT-Linear | 0.8730125079499682 | |
| BERT-Linear-CRF | 0.8774603174603175 | ||
| BERT-Prompt | 0.8549382716049383 |
Sampling 1% data in trainset by fixed random seed. They used the same hyper parameters in full data experiments.
Few Shot Test on CoNLL2003:
Sampled 69 samples(total 6973). Here list the number of entities in few shot dataset:
{'MISC': 51, 'ORG': 51, 'PER': 59, 'LOC': 90}| Model | Overall Span-Based F1 On Full Testset |
| bert-base-uncased | |
| BERT-Linear | 0.6778304852260387 |
| BERT-Linear-CRF | 0.6773130256876562 |
| BERT-Prompt | 0.7524185216492908 |
| BERT-BiLSTM-Linear | 0.037065541975802724 |
| BERT-BiLSTM-Linear-CRF | 0.029508301201363885 |
Few Shot Test on CCKS2019:
Sampled 10 samples(total 1000). Here list the numbers of entities in few shot dataset:
{'手术': 9, '影像检查': 5, '疾病和诊断': 45, '解剖部位': 48, '实验室检验': 19, '药物': 10}| Model | Overall Span-Based F1 On Full Testset | |
| bert-base-chinese | medbert-base-chinese | |
| BERT-Linear | 0.43918064570513354 | 0.47296831955922863 |
| BERT-Linear-CRF | 0.47901807928346324 | 0.537369759619329 |
| BERT-Prompt | 0.0038852361028093247 | 0.43338090840399623 |
Few Shot Test on NCBI-disease:
Sampled 54 samples(total about 5400) contains 41 disease entities.
| Model | Overall Span-Based F1 On Full Testset | |
| bert-base-uncased | medbert-base-chinese | |
| BERT-Linear | 0.6449976947902258 | 0.6617314414970182 |
| BERT-Linear-CRF | 0.6520612485276797 | 0.6777615976700645 |
| BERT-Prompt | 0.5552599758162031 | 0.5978593272171254 |