Skip to content

Latest commit

 

History

History
133 lines (108 loc) · 4.97 KB

README.md

File metadata and controls

133 lines (108 loc) · 4.97 KB
Raw

BatteryBERT

License

BatteryBERT: A Pre-trained Language Model for Battery Database Enhancement

Features

  • BatteryBERT Pre-training: Pre-training from Scratch or Further Training
  • BatteryBERT Fine-tuning: Sequence Classification + Question Answering
  • BatteryBERT Usage: Document Classifier, Device Data Extractor, General Q&A Agent
  • Large-scale Device Data Extraction and Database Enhancement

Installation

Run the following commands to clone the repository and install batterybert:

git clone https://github.com/ShuHuang/batterybert.git
cd batterybert; pip install -r requirements.txt; python setup.py develop

Usage

BatteryBERT Pre-training

Run pre-training:

Pre-training from scratch or further training using a masked language modeling (MLM) loss. See python run_pretraining.py --help for a full list of arguments and their defaults.

python run_pretrain.py \
    --train_root $TRAIN_ROOT \
    --eval_root $EVAL_ROOT \
    --output_dir $OUTPUT_DIR \
    --tokenizer_root $TOEKNIZER_ROOT \
    --checkpoint=$CHECKPOINT_DIR

Create a new WordPiece tokenizer:

Train a WordPiece tokenizer from scratch using the dataset from $TRAIN_ROOT. See python run_tokenizer.py --help for a full list of arguments and their defaults.

python run_tokenizer.py \
    --train_root $TRAIN_DIR \
    --save_root $SAVE_DIR \
    --save_name $TOKENIZER_NAME

BatteryBERT Fine-tuning

Run fine-tuning (question answering):

Fine-tune a BERT model on a question answering dataset (e.g. SQuAD). See python run_finetune_qa.py --help for a full list of arguments and their defaults.

python run_finetune_qa.py 
    --model_name_or_path $MODEL_NAME_OR_PATH \
    --output_dir $OUTPUT_DIR

Run fine-tuning (document classification):

Fine-tune a BERT model on a sequence classification dataset (e.g. paper corpus). See python run_finetune_doc_classify.py --help for a full list of arguments and their defaults.

$ python run_finetune_doc_classify.py 
    --model_name_or_path $MODEL_NAME_OR_PATH \
    --output_dir $OUTPUT_DIR \
    --train_root $TRAIN_ROOT \
    --eval_root $EVAL_ROOT

BatteryBERT Usage

Use the battery paper classifier:

>>> from batterybert.apps import DocClassifier

# Model name to be changed after published
# Create a binary classifier
>>> model_name = "batterydata/test4"
>>> sample_text = "sample text"
>>> classifier = DocClassifier(model_name)

# 0 for non-battery text, 1 for battery text
>>> category = classifier.classify(sample_text)
>>> print(category)

0

Use the device data extractor:

>>> from batterybert.apps import DeviceDataExtractor

# Model name to be changed after published
# Create a device data extractor
>>> model_name = "batterydata/test1"
>>> sample_text = "The anode of this Li-ion battery is graphite."
>>> extractor = DeviceDataExtractor(model_name)

# Set the confidence score threshold
>>> result = extractor.extract(sample_text, threshold=0.1)
>>> print(result)

[{'type': 'anode', 'answer': 'graphite', 'score': 0.9736555218696594, 'context': 'The anode of this battery is graphite.'}]

Use the general Q&A agent:

>>> from batterybert.apps import QAAgent

# Model name to be changed after published
# Create a QA agent
>>> model_name = "batterydata/test1"
>>> context = "The University of Cambridge is a collegiate research university in Cambridge, United Kingdom. Founded in 1209 and granted a royal charter by Henry III in 1231, Cambridge is the second-oldest university in the English-speaking world and the world's fourth-oldest surviving university."
>>> question = "When was University of Cambridge founded?"
>>> qa_agent = QAAgent(model_name)

# Set the confidence score threshold
>>> result = qa_agent.answer(question=question, context=context, threshold=0.1)
>>> print(result)

{'score': 0.9867061972618103, 'start': 105, 'end': 109, 'answer': '1209'}

Acknowledgements

This project was financially supported by the Science and Technology Facilities Council (STFC), the Royal Academy of Engineering (RCSRF1819\7\10) and Christ's College, Cambridge. The Argonne Leadership Computing Facility, which is a DOE Office of Science Facility, is also acknowledged for use of its research resources, under contract No. DEAC02-06CH11357.

Citation

@article{huang2022batterybert,
  title={BatteryBERT: A Pretrained Language Model for Battery Database Enhancement},
  author={Huang, Shu and Cole, Jacqueline M},
  journal={J. Chem. Inf. Model.},
  year={2022},
  doi={10.1021/acs.jcim.2c00035},
  url={DOI:10.1021/acs.jcim.2c00035},
  pages={DOI: 10.1021/acs.jcim.2c00035},
  publisher={ACS Publications}
}

DOI