🧪 ReactionMiner V2

This is the updated version of ReactionMiner. We have updated the pdf2txt module with Grobid + s2orc for better preprocessing. We have also updated the extraction module by finetuning Llama-3.1-8b with LoRA and vLLM implementation.

Official Repository for the EMNLP 2023 Demo Paper
Reaction Miner: An Integrated System for Chemical Reaction Extraction from Textual Data

🛠️ Environment

To get started, install the necessary packages:

pip install -r requirements.txt
python -m spacy download en_core_web_sm

For using the PDF-to-Text module in Reaction Miner, install the submodule:

git submodule update --init

Then follow the instructions from s2orc to install Grobid and relevant environment setup.

📖 How to Use Reaction Miner

Current updated version requires two different environments. The pdf2txt module is running under the environment described in s2orc. The rest is running under our environment.

Step 1: PDF-to-Text Conversion

Ensure Grobid server is running, then execute

bash pdf2text/pdf2txt.sh

This will process all the PDF files indicated by the defaulty directory defaultDir in pdf2text/config.py.

Given a PDF file, please refer to example.py to run the rest of our system. It can be broken down into the following 2 steps:

Step 2: Text Segmentation

Identifies paragraphs about chemical reactions and segments them:

from segmentation.segmentor import TopicSegmentor
segmentor = TopicSegmentor()
seg_texts = segmentor.segment(paragraphs)

Step 3: Reaction Extraction

Extracts structured chemical reactions from each segment:

from extraction.extractor import ReactionExtractor
extractor = ReactionExtractor('8b')
reactions = extractor.extract(seg_texts)

📚 Citation

If you find Reaction Miner helpful, please kindly cite our paper:

@inproceedings{zhong2023reaction,
  title={Reaction Miner: An Integrated System for Chemical Reaction Extraction from Textual Data},
  author={Zhong, Ming and Ouyang, Siru and Jiao, Yizhu and Kargupta, Priyanka and Luo, Leo and Shen, Yanzhen and Zhou, Bobby and Zhong, Xianrui and Liu, Xuan and Li, Hongxiang and others},
  booktitle={Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: System Demonstrations},
  pages={389--402},
  year={2023}
}

Running in Docker

Place any input files in ./inputs

To run Grobid + ReactionMiner using our pre-built image:

docker compose up -d

Output files will be located in ./results

When you're done using it, shut down related services using:

docker compose down

Running with GPU

By default, this image will not utilize a GPU if one is present

To use a GPU, uncomment the deploy section in docker-compose.yml

Building Docker Image

To build a Docker container for ReactionMiner:

docker compose build

You can build + run the image in one step using:

docker compose up -d --build

Running in Kubernetes

We have also included a kubernetes/ folder that contains experimental manifests for running in a Kubernetes cluster

You will need to adjust these manifests based on your own cluster.

First, create a Secret containing your HuggingFace API Token:

vi kubernetes/huggingface.secret.yml
kubectl apply -f kubernetes/huggingface.secret.yml

Warning: Do not commit this API token to source control

Next, start up Grobid:

kubectl apply -f kubernetes/grobid.yml

Finally, run a ReactionMiner Job:

kubectl apply -f kubernetes/reactionminer.job.yml

This will utilize your HuggingFace API Token secret and will make requests to the Grobid server.

To run an interactive shell with ReactionMiner for testing/debugging, you can use the Pod manifest:

kubectl apply -f kubernetes/reactionminer.pod.yml
kubectl exec -it reactionminer -- bash