MatSciBERT

A Materials Domain Language Model for Text Mining and Information Extraction

Using MatSciBERT

Installing requirements

bash install_requirements.sh

Importing libraries

import torch
from normalize_text import normalize
from transformers import AutoModel, AutoTokenizer

Loading pre-trained model and tokenizer

tokenizer = AutoTokenizer.from_pretrained('m3rg-iitd/matscibert')
model = AutoModel.from_pretrained('m3rg-iitd/matscibert')

Tokenizing sentences

sentences = ['SiO2 is a network former.']
norm_sents = [normalize(s) for s in sentences]
tokenized_sents = tokenizer(norm_sents)
tokenized_sents = {k: torch.Tensor(v).long() for k, v in tokenized_sents.items()}

Obtaining MatSciBERT embeddings

with torch.no_grad():
    last_hidden_state = model(**tokenized_sents)[0]

Citing

If you use MatSciBERT in your research, please cite MatSciBERT: A materials domain language model for text mining and information extraction

@article{gupta_matscibert_2022,
  title   = "{MatSciBERT}: A Materials Domain Language Model for Text Mining and Information Extraction",
  author  = "Gupta, Tanishq and 
            Zaki, Mohd and 
            Krishnan, N. M. Anoop and 
            Mausam",
  year    = "2022",
  month   = may,
  journal = "npj Computational Materials",
  volume  = "8",
  number  = "1",
  pages   = "102",
  issn    = "2057-3960",
  url     = "https://www.nature.com/articles/s41524-022-00784-w",
  doi     = "10.1038/s41524-022-00784-w"
}

Fine-tuned models for Named entity recognition, Abstract classification, and Relation classification with the codes of this repository are available on