Auto-CORPus

Requires python 3.6+

The Automated pipeline for Consistent Outputs from Research Publications (Auto-CORPus) is a tool for the standardisation and conversion of publication HTML to three convenient machine-interpretable outputs to support biomedical text analytics. Firstly, Auto-CORPus can be configured to convert HTML from various publication sources to BioC format. Secondly, Auto-CORPus transforms publication tables to a JSON format to store, exchange and annotate table data between text analytics systems. Finally, Auto-CORPus extracts abbreviations declared within publication text and provides an abbreviations JSON output that relates an abbreviation with the full definition.

We present a JSON format for sharing table content and metadata that is based on the BioC format. The JSON schema for the tables JSON can be found within the keyfiles directory.

Config files

If you wish to contribute or edit a config file then please follow the instructions in the config guide

Auto-CORPus is able to parse HTML from different publishers, which utilise different HTML structures and naming conventions. This is made possible by the inclusion of config files which tell Auto-CORPus how to identify specific sections of the article/table within the source HTML. We have supplied a config template along with example config files for PubMed Central, Plos Genetics and Nature Genetics in the configs directory. Documentation on how to create and modify config files is available within the Tutorial directory. Users of Auto-CORPus can submit their own config files for different sources via the issues tab.

Auto-CORPus recognises 3 categories of input file which are:

• Full text HTML documents covering the entire article
• HTML files which describe a single table
• Supplementary files (PDF, Word, Spreadsheets) within a folder named e.g PMC123_supplementary

Current work in progress is extending this to include images of tables. See the Alpha Testing section below.

Auto-CORPus does not provide functionality to retrieve input files directly from the publisher. Input file retrieval must be completed by the user in a way which the publisher permits.

Auto-CORPus relies on a standard naming convention to recognise the files and identify the correct order of tables. The naming convention can be seen below:

Full article HTML: {any_name_you_want}.html

• {any_name_you_want} is how Auto-CORPus will group articles and linked tables/image files

Linked table HTML: {any_name_you_want}_table_X.html

• {any_name_you_want} must be identical to the name given to the full text file followed by _table_X where X is the table number

If passing a single file via the file path then that file will be processed in the most suitable manner, if a directory is passed then Auto-CORPus will first group files based on common elements in their file name {any_name_you_want} and process all related files at once. Related files in separate directories will not be processed at the same time. Files processed at the same time will be output into the same files, an example input and output directory can be seen below:

Input:

PMC1.html
PMC1_table_1.html
PMC1_table_2.html
/subdir
    PMC1_table_3.html
    PMC1_table_4.html
/PMC1_supplementary
    /PMC1_data_fileA.docx
    /PMC1_data_fileB.pdf
    /PMC1_data_fileC.xlsx

Output:

PMC1_bioc.json
PMC1_abbreviations.json
PMC1_tables.json (contains table 1 & 2 and any tables described within the main text)
/subdir
    PMC1_tables.json (contains tables 3 & 4 only)
/PMC1_supplementary
    /PMC1_data_fileA_bioc.docx.json
    /PMC1_data_fileB_bioc.pdf.json
    /PMC1_data_fileC_bioc.xlsx.json
    /PMC1_data_fileA_tables_bioc.docx.json
    /PMC1_data_fileB_tables_bioc.pdf.json
    /PMC1_data_fileC_tables_bioc.xlsx.json

A log file is produced in the output directory providing details of the day/time Auto-CORPus was run, the arguments used and information about which files were successfully/unsuccessfully processed with a relevant error message.

Getting started:

Clone the repo, e.g.:

$ git clone [email protected]:omicsNLP/Auto-CORPus.git or (using HTTPS) git clone https://github.com/omicsNLP/Auto-CORPus.git

$ cd Auto-CORPus

$ python3 -m venv env or (for Windows users) py -[v] -m venv env (where v is the version of Python used)

$ source env/bin/activate or (for Windows users) path/to/env/Scripts/activate.bat

$ pip install .

You might get an error here ModuleNotFoundError: No module named 'skbuild' if you do then run

$ pip install --upgrade pip

Or you might need to install the Microsoft Build Tools for Visual Studio (see https://www.scivision.dev/python-windows-visual-c-14-required for minimal installation requirements so that python-Levenshtein package can be installed) first and then re-run

$ pip install .

Conda installation

$ conda create -n autocorpus python==3.9 pip python-levenshtein==0.12.2

$ conda activate autocorpus

$ pip install .

Try the install

Run the below command for a single file example

$ python run_app.py -c "configs/config_pmc.json" -t "output" -f "path/to/html/file" -o JSON

Run the below command for a directory of files example

$ python run_app.py -c "configs/config_pmc.json" -t "output" -f "path/to/directory/of/html/files" -o JSON

Available arguments:

-f (input file path) - file or directory to run Auto-CORPus on

-t (output file path) - file path where Auto-CORPus should output files

-c (config) - which config file to use

-o(output format) - either JSON or XML (defaults to JSON)

Alpha testing

We are developing an Auto-CORPus plugin to process images of tables and we include an alpha version of this functionality. Table image files can be processed in either .png or .jpeg/jpg formats. We are working on improving the accuracy of both the table layout and character recognition aspects, and we will update this repo as the plugin advances.

We utilise opencv for cell detection and tesseract for optical character recognition. Tesseract will need to be installed separately onto your system for the table image recognition aspect of Auto-CORPus to work. Please follow the guidance given by tesseract on how to do this.

We have made trained datasets available for use with this feature, but we will continue to train these datasets to increase their accuracy, and it is very likely that the trained datasets we offer will be updated frequently during active development periods.

As with HTML input files, the image input files should be retrieved by the user in a way which the publisher permits. The naming convention is:

Table image file: {any_name_you_want}_table_X.png/jpg/jpeg

• {any_name_you_want} must be identical to the name given to the full text file followed by _table_X where X is the table number

Additional argument:

-s (trained dataset) - trained dataset to use for pytesseract OCR. Value should be given in a format recognised by pytesseract with a "+" between each datafile, such as "eng+all".