scibart

So far, this is a refactoring job of a notebook in CurationCorp's amazing curation-corpus repository for training on GPU clusters to tune BART for abstractive summarization of scientific literature.

Part of the Coronawhy project.

How to create a dataset from scratch

Currently, the dataset is sourced as follows:

• Text-abstract pairs from Arxiv and the Semantic Scholar Corpus as provided by Santosh-Gupta's ScientificSummarizationDataSets repo
• Text-headline pairs from WikiHow, provided by mahnazkoupaee's WikiHow-Dataset repo
• Curation Corpus

To create a new dataset from scratch:

1. Download the ArXiv and Semantic Scholar Corpus datasets from gdrive (as described here) and unzip into raw_data/ArxivStructuredAbstractSectionalSummaries and raw_data/SemanticScholarAbstractSectionSummaryDataSet
2. Download wikihowAll.csv (as described here) into raw_data/wikihow
3. Scrape the Curation Corpus dataset as explained in the repo, then move curation-corpus-base-with-articles.csv to raw_data/curation_corpus
4. Run python src/data/create_dataset.py. This will create a new folder called data with ~40 compressed parquet files

The current dataset is stored in a single pandas dataframe with the following schema:

Column name	Column Type	Description
text	str	Original text on which the summary is based
summary	str	Summary of the original text
data_src	str	Directory name of the original dataset in raw_data