ArxivDIGESTables

The data set of tables released with ArxivDIGESTables: Synthesizing Scientific Literature into Tables using Language Models.

The tables in this dataset are arbitrary json objects, so they don't play nicely with huggingface's load_dataset method. The recommended way to access the data is to download individual files from huggingface.co/datasets/blnewman/arxivDIGESTables.

For the high quality data, you should download papers.jsonl, tables.jsonl, and full_texts.jsonl.gz. If you want more tables that are less stringently filtered and do not have associated full texts, you can download papers_medium_quality.jsonl and tables_medium_quality.jsonl.

• The papers*.jsonl files include information about the papers cited in the tables including their S2 corpus ids, title, abstract and the ids of what tables they can be found in. These are mostly useful for analysis as almost all of the information is also included in the tables*.jsonl files.
• The tables*.jsonl files include:
  • tabid: an id for each table
  • table: the table itself, which is a nested json dictionary
  • row_bib_map: which maps each row of the table to the corpus id, title, and abstract for the paper cited in that row.
  • caption: the table's caption
  • in_text_ref: a list of paragraphs where the table is refered to in the main text
  • arxiv_id: the arxiv id of the paper that table comes from
• full_text.jsonl.gzcontains the full texts for the papers in papers.jsonl.

If you want to preview the tables, you can use huggingface dataset's loader. In this case, the tables are stored as json strings and need to be parsed:

import json
from datasets import load_dataset

# high quality
tables = load_dataset("blnewman/arxivDIGESTables")

# load the table from json string. Not necessary if you download `tables.jsonl` directly.
print(json.loads(tables["validation"]["table"][0]))

# medium quality
arxivdigestables_medium = load_dataset("blnewman/arxivDIGESTables", "medium_quality")

See the github repo for more details.

Citation

@article{newman2024arxivdigestables,
      title={ArxivDIGESTables: Synthesizing Scientific Literature into Tables using Language Models}, 
      author={Benjamin Newman and Yoonjoo Lee and Aakanksha Naik and Pao Siangliulue and Raymond Fok and Juho Kim and Daniel S. Weld and Joseph Chee Chang and Kyle Lo},
      year={2024},
      journal={arXiv preprint},
      url={https://arxiv.org/abs/2410.22360}, 
}