My public talk about this project at Sberloga:
QaNER - NER via Exractive QA

Unofficial implementation of QaNER: Prompting Question Answering Models for Few-shot Named Entity Recognition.

You can adopt this pipeline for arbitrary BIO-markup data.

pip install qaner

Pipeline results on CoNLL-2003 dataset:

• Metrics
• Trained Hugging Face model

Script for training QaNER model:

qaner-train \
--bert_model_name 'bert-base-uncased' \
--path_to_prompt_mapper 'data/conll2003/prompt_mapper.json' \
--path_to_train_data 'data/conll2003/train.bio' \
--path_to_test_data 'data/conll2003/test.bio' \
--path_to_save_model 'dayyass/qaner-conll-bert-base-uncased' \
--n_epochs 2 \
--batch_size 128 \
--learning_rate 1e-5 \
--seed 42 \
--log_dir 'runs/qaner'

Required arguments:

• --bert_model_name - base bert model for QaNER fine-tuning
• --path_to_prompt_mapper - path to prompt mapper json file
• --path_to_train_data - path to train data (BIO-markup)
• --path_to_test_data - path to test data (BIO-markup)
• --path_to_save_model - path to save trained QaNER model
• --n_epochs - number of epochs to fine-tune
• --batch_size - batch size
• --learning_rate - learning rate

Optional arguments:

• --seed - random seed for reproducibility (default: 42)
• --log_dir - tensorboard log_dir (default: 'runs/qaner')

Script for inference trained QaNER model:

qaner-inference \
--context 'EU rejects German call to boycott British lamb .' \
--question 'What is the organization?' \
--path_to_prompt_mapper 'data/conll2003/prompt_mapper.json' \
--path_to_trained_model 'dayyass/qaner-conll-bert-base-uncased' \
--n_best_size 1 \
--max_answer_length 100 \
--seed 42

Result:

question: What is the organization?

context: EU rejects German call to boycott British lamb .

answer: [Span(token='EU', label='ORG', start_context_char_pos=0, end_context_char_pos=2)]

Required arguments:

• --context - sentence to extract entities from
• --question - question prompt with entity name to extract (examples below)
• --path_to_prompt_mapper - path to prompt mapper json file
• --path_to_trained_model - path to trained QaNER model
• --n_best_size - number of best QA answers to consider

Optional arguments:

• --max_answer_length - entity max length to eliminate very long entities (default: 100)
• --seed - random seed for reproducibility (default: 42)

Possible inference questions for CoNLL-2003:

• What is the location? (LOC)
• What is the person? (PER)
• What is the organization? (ORG)
• What is the miscellaneous entity? (MISC)

Python >= 3.7

@misc{liu2022qaner,
    title         = {QaNER: Prompting Question Answering Models for Few-shot Named Entity Recognition},
    author        = {Andy T. Liu and Wei Xiao and Henghui Zhu and Dejiao Zhang and Shang-Wen Li and Andrew Arnold},
    year          = {2022},
    eprint        = {2203.01543},
    archivePrefix = {arXiv},
    primaryClass  = {cs.LG}
}