markdown-body entry-content container-lg" itemprop="text">

A Parser That Adds Special Identifiers To Markdown Files For Deep Learning

GitHub contains a large corpus data that is amenable for NLP, in the form of Issues, READMEs, pull request comments and other items. However, this text is often accompanies by markdown which allows the user to specify styling (bold, underline, headings) and specialized formatting (code blocks, tables, block quotes, hyperlinks). This library has two goals:

1. Insert custom field indicators

This is so markdown information is not lost. For example, a list block is enclosed with xxxlistB and xxxlistE and a code block is enclosed with xxxcdb and xxxcde. Other noteable examples:

@mentions: xxxatmention (the handle is removed and replaced by just this indicator)
quote blocks: xxxqb/xxxqe
strikethrough: xxxdelb/xxxdele
horizontal rule: xxxhr
{large, medium, small} headers: annotated with xxxh{l,m,s}. H1=large, H2-3=medium, H4-6=small.

2. Discard superfluous information

GitHub issues often contain a large stack trace, or a large table with data. This library comes equipped with sensible defaults to surface the most relevant information and discard what would otherwhise be lots of characters for a machine learning algorithm to handle:

Code Blocks: only first two and last two rows are kept
Tables: only table headers are kept
Urls: only the host is kept. For example www.google.com/search is reformatted to www.google.com
Images: the image is discarded but the file extension and metadata about the image (available to screenreader) is extracted.
IP Addresses, extremely long numbers are marked as xxunk

This parser works by converting markdown to HTML then converting the HTML (along with some of the HTML tags, in certain cases) to text.

Installation

pip install mdparse

Caveats

This library makes extremely opinionated choices on how to parse markdown and filter information. This library is for experimental purposes only, and may not be appropriate for every problem. Please use with caution.

The primary use case of this parser has been to prepare a large corpus of GitHub Issue data for a language model. However, we envision this parser would be applicable to other machine learning tasks involving the extraction of features from the text of GitHub Issues, Readme files, or pull request comments.

Examples

See /notebooks/Demo.ipynb for an example of the transformations this parser does on a markdown file.

License

machine-learning-apps/mdparse

ErrorLooks like something went wrong!

About

Topics

Resources

License

Uh oh!

Stars

Watchers

Forks

Releases 3

Packages 0

Uh oh!

Languages

Error
Looks like something went wrong!

Packages