Detecting claim from scientific publication using discourse model and transfer learning. Models are trained using AllenNLP library.

You can install the package using PIP, which will help you use the discourse classes inside a module

pip install git+https://github.com/titipata/detecting-scientific-claim.git

you will be able to use them as

import discourse
predictor = discourse.DiscourseCRFClassifierPredictor()

Running AllenNLP to train a discourse model using PubmMedRCT dataset as follows

allennlp train experiments/pubmed_rct.json -s output --include-package discourse

We point data location to Amazon S3 directly in pubmed_rct.json so you do not need to download the data locally. Change cuda_device to -1 in pubmed_rct.json if you want to run on CPU. There are more experiments available in experiments folder.

Note that you have to remove output folder first before running.

We trained the Bidirectional LSTM model on structured abstracts from Pubmed to predict discourse probability (RESULTS, METHODS, CONCLUSIONS, BACKGROUND, OBJECTIVE) of a given sentence. You can download trained model from Amazon S3

wget https://s3-us-west-2.amazonaws.com/pubmed-rct/model.tar.gz # or model_crf.tar.gz for pretrained model with CRF layer

and run web service for discourse prediction task as follow

bash web_service.sh

To test the train model with provided examples fixtures.json, simply run the following to predict labels.

allennlp predict model.tar.gz \
    pubmed-rct/PubMed_200k_RCT/fixtures.json \
    --include-package discourse \
    --predictor discourse_predictor

or run the following for

allennlp predict model_crf.tar.gz \
    pubmed-rct/PubMed_200k_RCT/fixtures_crf.json \
    --include-package discourse \
    --predictor discourse_crf_predictor

To evaluate discourse model, you can run the following command

allennlp evaluate model.tar.gz \
  https://s3-us-west-2.amazonaws.com/pubmed-rct/test.json \
  --include-package discourse

We use transfer learning with fine tuning to train claim extraction model from pre-trained discourse model. The schematic of the training can be seen below.

You can run the demo web application to detect claims as follows

export FLASK_APP=main.py
flask run --host=0.0.0.0 # this will serve at port 5000

The interface will look something like this

And output will look something like the following (highlight means claim, tag behind the sentence is discourse prediction)

Expertly annotated dataset We release the dataset of annotated 1,500 abstracts containing 11,702 sentences (2,276 annotated as claim sentences) sampled from 110 biomedical journals. The final dataset are the majority vote from three experts. The annotations are hosted on Amazon S3 and can be found from these given URLs.

• Python 3.6
• AllenNLP >= 0.6.1
• spacy
• fastText
• Pubmed RCT - dataset

You can cite our paper available on arXiv as

Achakulvisut, Titipat, Chandra Bhagavatula, Daniel Acuna, and Konrad Kording. "Claim Extraction in Biomedical Publications using Deep Discourse Model and Transfer Learning." arXiv preprint arXiv:1907.00962 (2019).

or using BibTeX

@article{achakulvisut2019claim,
  title={Claim Extraction in Biomedical Publications using Deep Discourse Model and Transfer Learning},
  author={Achakulvisut, Titipat and Bhagavatula, Chandra and Acuna, Daniel and Kording, Konrad},
  journal={arXiv preprint arXiv:1907.00962},
  year={2019}
}

This project is done at the Allen Institute for Artificial Intelligence and Konrad Kording lab, University of Pennsylvania