SciBERT

SciBERT is a BERT model trained on scientific text.

• SciBERT is trained on papers from the corpus of semanticscholar.org. Corpus size is 1.14M papers, 3.1B tokens. We use the full text of the papers in training, not just abstracts.

• SciBERT has its own vocabulary (scivocab) that's built to best match the training corpus. We trained cased and uncased versions. We also include models trained on the original BERT vocabulary (basevocab) for comparison.

• It results in state-of-the-art performance on a wide range of scientific domain nlp tasks. The details of the evaluation are in the paper. Evaluation code and data are included in this repo.

Downloading Trained Models

Update! SciBERT models now installable directly within Huggingface's framework under the allenai org:

from transformers import *

tokenizer = AutoTokenizer.from_pretrained('allenai/scibert_scivocab_uncased')
model = AutoModel.from_pretrained('allenai/scibert_scivocab_uncased')

tokenizer = AutoTokenizer.from_pretrained('allenai/scibert_scivocab_cased')
model = AutoModel.from_pretrained('allenai/scibert_scivocab_cased')

We release the tensorflow and the pytorch version of the trained models. The tensorflow version is compatible with code that works with the model from Google Research. The pytorch version is created using the Hugging Face library, and this repo shows how to use it in AllenNLP. All combinations of scivocab and basevocab, cased and uncased models are available below. Our evaluation shows that scivocab-uncased usually gives the best results.

Tensorflow Models

• scibert-scivocab-uncased (Recommended)
• scibert-scivocab-cased
• scibert-basevocab-uncased
• scibert-basevocab-cased

PyTorch AllenNLP Models

• scibert-scivocab-uncased (Recommended)
• scibert-scivocab-cased
• scibert-basevocab-uncased
• scibert-basevocab-cased

PyTorch HuggingFace Models

• scibert-scivocab-uncased (Recommended)
• scibert-scivocab-cased
• scibert-basevocab-uncased
• scibert-basevocab-cased

Using SciBERT in your own model

SciBERT models include all necessary files to be plugged in your own model and are in same format as BERT. If you are using Tensorflow, refer to Google's BERT repo and if you use PyTorch, refer to Hugging Face's repo where detailed instructions on using BERT models are provided.

Training new models using AllenNLP

To run experiments on different tasks and reproduce our results in the paper, you need to first setup the Python 3.6 environment:

pip install -r requirements.txt

which will install dependencies like AllenNLP.

Use the scibert/scripts/train_allennlp_local.sh script as an example of how to run an experiment (you'll need to modify paths and variable names like TASK and DATASET).

We include a broad set of scientific nlp datasets under the data/ directory across the following tasks. Each task has a sub-directory of available datasets.

├── ner
│   ├── JNLPBA
│   ├── NCBI-disease
│   ├── bc5cdr
│   └── sciie
├── parsing
│   └── genia
├── pico
│   └── ebmnlp
└── text_classification
    ├── chemprot
    ├── citation_intent
    ├── mag
    ├── rct-20k
    ├── sci-cite
    └── sciie-relation-extraction

For example to run the model on the Named Entity Recognition (NER) task and on the BC5CDR dataset (BioCreative V CDR), modify the scibert/train_allennlp_local.sh script according to:

DATASET='bc5cdr'
TASK='ner'
...

Decompress the PyTorch model that you downloaded using
tar -xvf scibert_scivocab_uncased.tar
The results will be in the scibert_scivocab_uncased directory containing two files: A vocabulary file (vocab.txt) and a weights file (weights.tar.gz). Copy the files to your desired location and then set correct paths for BERT_WEIGHTS and BERT_VOCAB in the script:

export BERT_VOCAB=path-to/scibert_scivocab_uncased.vocab
export BERT_WEIGHTS=path-to/scibert_scivocab_uncased.tar.gz

Finally run the script:

./scibert/scripts/train_allennlp_local.sh [serialization-directory]

Where [serialization-directory] is the path to an output directory where the model files will be stored.

Citing

If you use SciBERT in your research, please cite SciBERT: Pretrained Language Model for Scientific Text.

@inproceedings{Beltagy2019SciBERT,
  title={SciBERT: Pretrained Language Model for Scientific Text},
  author={Iz Beltagy and Kyle Lo and Arman Cohan},
  year={2019},
  booktitle={EMNLP},
  Eprint={arXiv:1903.10676}
}

SciBERT is an open-source project developed by the Allen Institute for Artificial Intelligence (AI2). AI2 is a non-profit institute with the mission to contribute to humanity through high-impact AI research and engineering.