KIST-CSRC / CO2RR_NER

Named entity recognition to CO2RR catalyst papers

CO2RR Named Entity Recognition (NER)

Introduction

1) Model structures for named entity recognition (NER)

2) Exclusion of noise papers using Doc2vec and latent dirichlet allocation (LDA)

Two techniques are needed to automatically extract data from the literature and utilize the extracted data. The first is a technology to selectively classify only the papers we want through preprocess of paper, and the second is a technology, called NER, to automatically extract information from the literature. Therefore, This repository contains how to exclude noise paper from corpus (paper preprocess) and CO2 reduction reaction (CO2RR) NER models (BERT, Bi-LSTM) that classify words in literature into predefined classes, such as catalyst and product.

1. Paper preprocess

1) Installation

Using conda

conda env create -f conda_requirements_paper.yaml

Using pip

pip install -r requirements_paper.txt

2) Download data files

git clone https://github.com/KIST-CSRC/CO2RR_NER.git
git lfs pull

3) Code usage

Examples: Train Doc2vec

from gensim.models.doc2vec import Doc2Vec,TaggedDocument
from chemdataextractor.doc.text import Text
from chemdataextractor.nlp.tokenize import ChemWordTokenizer
import os
import time  
import numpy as np
from Paper_preprocess.Doc2vec import DocumentVector

path_dir = 'your_corpus_path'
documentvector = DocumentVector()
documentvector.run(path_dir)
print(time.strftime('%c', time.localtime(time.time())))  

Examples: Train LDA

from gensim.models.ldamodel import LdaModel
from chemdataextractor.doc.text import Text
from chemdataextractor.nlp.tokenize import ChemWordTokenizer
import os
import time  
import pickle

from nltk.corpus import stopwords 
from gensim import corpora
import pandas as pd
import csv
import pyLDAvis
import pyLDAvis.gensim_models

from gensim.models.coherencemodel import CoherenceModel 
import matplotlib.pyplot as plt
import warnings

import nltk
nltk.download('averaged_perceptron_tagger')  
from Paper_preprocess.Lda import Lda

path_dir = 'your_corpus_path'
lda_obj = Lda()
lda_obj.run(path_dir)

print(time.strftime('%c', time.localtime(time.time())))

Examples: Calculate cosine similarity

1. Load the trained doc2vec model into your corpus.
2. Set seed papers that are examples of papers you want to selectively collect.
3. Calculate the cosine similarity of each paper using Calculate_cosine_similarity.py

Finally, the noise paper can be removed using the calculated cosine similarity and lda result. Please refer to this paper for details

2. CO2RR NER (BERT_CRF)

1) Installation

Using conda

conda env create -f conda_requirements_ner.yaml

Using pip

pip install -r requirements_ner.txt

2) Download data files

git clone https://github.com/KIST-CSRC/CO2RR_NER.git
git lfs pull

The MatBERT model retrained with CO2RR was large, so it was uploaded using large file storage (LFS). Therefore, the "git lfs pull" command is absolutely necessary.

3) Pre-trained Model

Please ensure that after set up the directories as follows:

BERT_CRF
├── best_model
│   └── .gitattributes
│   └── best.pt
├── dataset
│   └── examples of data (such as new_IOBES_bert_base_cased.csv)
├── models
│   └── base_ner_model.py
│   └── bert_model.py
│   └── crf.py
├── predict
│   └── dataset
│       └── ACS_CO2RR_00343.csv
│       └── ACS_CO2RR_00343.json
├── utils
│   └── data.py
│   └── metrics.py
│   └── post_process.py
└── predict.py
└── run.py

4) Code usage

Model pretrainig and finetuning is done by modifying the code opensourced by MatBERT_NER authors to suit our work.

Examples: Re-train BERT model

The code below is part of the run.py file. After modifying the parameters of the corresponding part as you want, if you run the run.py file, learning will proceed.

n_epochs = 100
full_finetuning = True
savename = time.strftime("%Y%m%d")

device = "cuda:1"
models = {'bert_base': 'bert-base-cased',
         'scibert': 'allenai/scibert_scivocab_cased',
         'matbert': './matbert-base-cased',
         'matscibert' : 'm3rg-iitd/matscibert',
         'bert_large' :'bert-large-cased'}

splits = {'_80_10_10': [0.8, 0.1, 0.1]}


label_type = 'BERT'
tag_type = 'IOBES'
# 0: exclude structure entities.
structure = '1'
batch_size = 32
k_fold = 10
datafile = f"./dataset/new_{tag_type}_matbert_cased.csv"

for alias, split in splits.items():
  # If you want to change the model you want to retrain, edit this part
   for model_name in ['matbert']:

Examples: Predicting samples

1. Like a file in the "predict/dataset/.." path, prepare two data files in csv and json format.
2. Run predict.py

CO2RR NER (Bi-LSTM CRF)

We also used the Bi-LSTM CRF as the NER model and compared the performance of the BERT CRF model. For more details, see the paper below. Download the Bi-LSTM CRF code here.

Reference

For more details, see the paper below. Please cite us if you are using our model in your research work:

[1] Choi, Jiwoo, et al. "Deep learning of electrochemical CO 2 conversion literature reveals research trends and directions." Journal of Materials Chemistry A (2023).