Automated Lit Review

Jason Portenoy 2018

This is code and sample data accompanying the paper:

Supervised Learning for Automated Literature Review

published in the proceedings of the 4th Joint Workshop on Bibliometric-enhanced Information Retrieval and Natural Language Processing for Digital Libraries (BIRNDL 2019) co-located with the 42nd International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR 2019)

Starting with a list of seed papers, get candidate papers by following in- and out-citations (2 degrees). Then, train a classifier to rank the candidate papers. Repeat this a number of times to get an aggregate ranking for many candidate papers.

Example script in scripts/run_autoreview.py

Inputs:

• List of paper IDs for the seed set.
• Data for paper citations.
• Paper data to be used as features for the classifiers (e.g., clusters, eigenfactor, titles, etc.)

Parameters:

• Size of the initial split
• Number of times to perform the overall process of collecting candidate papers and training a classifier

Output:

• List of papers not in the seed set, ordered descending by relevance score.

Installation

Install via PyPI:

pip install autoreview

Example

• Apache Spark (https://spark.apache.org/downloads.html) must be installed to run the example.

• The environment variable SPARK_HOME must be set (preferably in a .env file) with the path to Spark.

  • Java version 8 is required to be used with Spark. Make sure Java 8 is installed and point to its path with the environment variable JAVA_HOME.

  • Example .env file:

    SPARK_HOME=/home/spark-2.4.0-bin-hadoop2.7
    JAVA_HOME=/usr/lib/jvm/java-8-openjdk-amd64
    

• Create a virtual environment and install the required libraries:

  virtualenv venv
  source venv/bin/activate
  pip install -r requirements.txt
  

• Run the full autoreview pipeline using sample data:

  python scripts/run_autoreview.py --id-list sample_data/sample_IDs_MAG.txt --citations sample_data/MAG_citations_sample --papers sample_data/MAG_papers_sample --sample-size 15 --random-seed 999 --id-colname Paper_ID --cited-colname Paper_Reference_ID --outdir sample_data/sample_output --debug
  

• This is just meant to show how the system operates. It will not provide meaningful results with such a small sample of paper and citation data.

• It will output the top predictions in sample_data/sample_output/predictions.tsv.

Development

For new releases:

# increment the version number
bump2version patch

Replace patch with minor or major as needed. Then:

# push new release to github
git push --tags

# build and upload to PyPI
python setup.py sdist bdist_wheel
twine check dist/*
twine upload dist/*