Skip to content
This repository has been archived by the owner on Feb 4, 2022. It is now read-only.
/ reach Public archive

Wellcome tool to parse references scraped from policy documents using machine learning

License

Notifications You must be signed in to change notification settings

wellcometrust/reach

Repository files navigation

Wellcome Reach

Wellcome Reach is an open source service for discovering how research publications are cited in global policy documents, including those produced by policy organizations such as the WHO, MSF, and the UK government. Key parts of it include:

  1. Web scrapers for pulling PDF "policy documents" from policy organizations,
  2. A reference parser for extracting references from these documents,
  3. A task for sourcing publications from Europe PMC (EPMC),
  4. A task for matching policy document references to EPMC publications,
  5. An Airflow installation for automating the above tasks, and
  6. A web application for searching and retrieving data from the datasets produced above.

Wellcome Reach is written in Python and developed using docker-compose. It's deployed into Kubernetes.

Although parts of the Wellcome Reach have been in use at Wellcome since mid-2018, the project has only been open source since March 2019. Given these early days, please be patient as various parts of it are made accessible to external users. All issues and pull requests are welcome. Contributing guidelines can be found in CONTRIBUTING.md.

Development

Dependencies

To develop for this project, you will need:

  1. Python 3.6+, plus pip and virtualenv
  2. Docker and docker-compose
  3. AWS credentials with read/write S3 permissions to an S3 bucket of your choosing.

docker-compose

To bring up the development environment using docker:

  1. Set your AWS credentials into your environment. (AWS_ACCESS_KEY_ID and AWS_SECRET_ACCESS_KEY)
  2. Build and start the env with:
    make docker-build
    docker-compose up -d
    
  3. Verify it all came up with:
    docker-compose ps
    

Once up, you'll be able to access:

virtualenv

For local development outside of airflow or other services, use the project's virtualenv:

make virtualenv
source build/virtualenv/bin/activate

Testing

To run all tests for the project using the official Python version and other dependencies, run:

make docker-test

You can also run tests locally using the project's virtualenv, with

make test

or using the appropriate pytest command, as documented in Makefile.

Airflow

Wellcome Reach uses Apache Airflow to automate running its data pipelines. Specifically, we've broken down the batch pipeline into a series of dependent steps, all part of a Directed Acyclic Graph (DAG).

Running a task in airflow

It's quite common to want to run a single task in Airflow without having to click through in the UI, not least because all logging messages are then on the console. To do this, from top of the project directory:

  1. Bring up the stack with docker-compose as shown above, and
  2. Run the following command, substituting for DAG_NAME, TASK_NAME, and JSON_PARAMS:
    ./docker_exec.sh airflow test \
        ${DAG_NAME} ${TASK_NAME} \
        2018-11-02 -tp '${JSON_PARAMS}'
    

For developers inside Wellcome

Although not required, you can add Sentry reporting from your local dev environment to a localdev project inside Wellcome's sentry account by running:

```
eval $(./export_wellcome_env.py)
```

before running docker-compose up -d above.

Deployment

For production, a typical deployment uses:

  • a Kubernetes cluster that supports persistent volumes
  • a PostgreSQL or MySQL database for Airflow to use
  • a distributed storage service such as S3
  • an ElasticSearch cluster for searching documents

Evaluation

The evaluation results are stored as an output here. Broadly the evaluation works by comparing a gold set of results - a manually annotated dataset of all the publications that should be found in a sample of policy documents, against the publications Reach identified in the same sample of policy documents. The evaluation script is held in another private repo.

Further reading

Contributing

See the Contributing guidelines

About

Wellcome tool to parse references scraped from policy documents using machine learning

Resources

License

Stars

Watchers

Forks

Packages

No packages published