Currently, we have a library in the arXiv standard format for papers. However, the formats for biology and computer science papers differ.

I was able to locate the benchmark data used by grobid on Zenodo, referred to as bioRxiv, and in the PMC_sample_1943 s3 bucket. However, there are more than a thousand files in total.

GROBID also conducted a benchmark on the PLOS journal and eLife. More details can be found on the GROBID documentation.

Additional benchmarks can be found here.

Building a Biology Collection for Refstudio

Realistically, researchers will be working with approximately 50-60 articles per paper. Greg already has a collection of Machine Learning and Computer Science papers, which are mainly arXiv papers and thus easier to parse.

I wanted to mimic the workflow of a researcher other than a computer scientist. I began with a renowned journal, Nature, and specifically focused on biology-related field papers.

I looked at the most-cited recent papers, such as this one from 2019.

Thus, I created a collection of 25 papers from an actual publication in Nature. These 25 papers derive from well-known journals such as Nature, Science, Cell, etc.

Some Observations

I imported them to Refstudio. Some of them were 30 pages long or 11 MB in size, so it took quite some time to import them. Extracting the full text of these papers can be time-consuming.

I suggest initially only importing the header metadata to facilitate the process. The full text extraction of the PDFs, as conducted here, can be done later. This approach will prevent the user from having to wait excessively after the initial import.

How to Share these PDFs?

I can share the links to these PDFs in the repository. The team can then download them using wget. There might be licensing issues, so we may not distribute the papers directly but merely share the links.

Other Tools for Scientific Full Text PDF Extraction

There are template-based PDF extraction tools available:

Code: PDF Data Extractor
Paper: Template-based extraction

Uh oh!

Build a library of PDFs that are hard to parse #111

Description

Metadata

Metadata

Assignees

Labels

Type

Fields

Projects

Milestone

Relationships

Development

Issue actions