S2ORC: The Semantic Scholar Open Research Corpus

S2ORC is a general-purpose corpus for NLP and text mining research over scientific papers.

• We've curated a unified resource that combines aspects of citation graphs (i.e. rich paper metadata, abstracts, citation edges) with a full text corpus that preserves important scientific paper structure (i.e. sections, inline citation mentions, references to tables and figures).
• Our corpus covers 136M+ paper nodes with 12.7M+ full text papers and connected by 467M+ citation edges by unifying data from many different sources covering many different academic disciplines and identifying open-access papers using services like Unpaywall.

For more details, see our ACL 2020 paper or watch our 12 min ACL 2020 talk.

For an example snippet of data, see the data/ directory in this repo.

• S2ORC is jointly maintained by Kyle Lo and Lucy Lu Wang at the Allen Institute for AI.
• S2ORC is only for non-commercial use, and is released under the CC BY-NC 2.0. By using S2ORC, you agree to the terms in the license.
• Feel free to email us and subscribe to our mailing list.
• Please cite our paper if you use S2ORC for your project. See the BibTeX.

News

• Feb 2021 - Released s2orc-doc2json to support parsing of PDF and LaTeX to JSON format.

Project status

We just released a new version, so are currently getting feedback from users, and supporting use of the corpus internally/externally for other projects. There are a few things on our TODO list.

Short-term

1. Refactoring LaTeX parser & doing a LaTeX release. This includes special handling of things we didn't consider before, like nested section headers and representing bullet-lists as paragraphs.

2. Adding other joined metadata, like arXiv categories.

3. Releasing the Grobid and LaTeX parsers publicly. - Released s2orc-doc2json on Feb 2021

4. Updating the logic around paper de-duplication to prioritize ACL draft over arXiv preprint draft, when possible.

5. Integrating PMC XML parses as well.

Medium-term (some work has started, but too early to tell)

1. Graph traversal in S2ORC is difficult due to its size. We're currently thinking of ways to make it easier to traverse the citation graph.

2. We're investigating computer-vision tools to clean up the Grobid parser output.

3. A PDF-to-S2ORC JSON service. - See s2orc-doc2json

Longer-term (thinking about it, but won't be for a while)

1. We have extracted images and would also like to release these with S2ORC, but licensing/copyright around images is a bit tricky (they're separately licensed from the papers themselves).

S2ORC releases

Release: 2020-07-05

• Released a new version of S2ORC containing papers up until 2020-04-14, bringing full text coverage from 8M to 12M.
• Lifted some paper filters to be more lenient toward papers that don't have sufficient amount of text. This brought total paper count to 136M from 81M.
• Updated the schema to keep paper metadata and parsed paper text separate.
• Fixed major bugs such as (i) missing section names, (ii) inline citation mention links that don't resolve to bibliographies, and (iii) unpredictable typing in certain metadata fields.
• Omitted LaTeX parses from this release. They will be added in a subsequent release. Part of the dataset schema change is to accommodate incremental releases (e.g. LaTeX-only release without having to re-run PDF parsing).

Status: 2020-04-07

• S2ORC has been accepted to ACL 2020!
• We've changed the name of the project to S2ORC. We will update the preprint shortly with the new name.
• The BibTeX citation has also been changed to reflect this.

Release: 2019-09-28

• Statistics: 81M+ paper nodes; 73M+ gold abstracts; 8M+ full text papers
• Due to release bugs (e.g. missing section names), we no longer recommend usage of this version. If you must use this version and need assistance, please contact Kyle and Lucy.
• MAG fields of study:

Download instructions

Please request access to S2ORC through this short Google Form and we'll send you URLs to download the dataset. It may take us up to a week to get back to you. If it has been longer than one week since you have completed the form and you have not heard from us, please complete the form again or send us an email -- your request may have slipped through the cracks.

Please note that S2ORC is only available for non-commercial use. Please make sure to see the CC BY-NC 2.0 for the terms of usage.

Valid release versions

• 20200705v1 (latest)
• 20190928 (no longer providing access, but still can help if needed)

Release directory structure

As of the latest release, the full corpus consists of metadata/ and pdf_parses/ splits, each containing 100 gzipped files. The directory structure for each release is:

yyyymmddvi/                 # version ID.  e.g. `20200705v1` means the first version
                            # released on 2020-07-05
|-- full/ 
   |-- metadata/            # 100 jsonl files; each file approx 529MB & 2.0GB unzipped
       |-- metadata_0.jsonl.gz
       |-- metadata_1.jsonl.gz
       ...
   |-- pdf_parses/          # 100 jsonl files; each file approx 1.6GB zipped & 6.5GB unzipped
       |-- pdf_parses_0.jsonl.gz
       |-- pdf_parses_1.jsonl.gz
       ...
|-- sample/
   |-- metadata/            # 1 jsonl file; first 1000 rows of full/metadata/metadata_0.jsonl
       |-- sample.jsonl
   |-- pdf_parses/          # 1 jsonl file; first 203 rows of full/pdf_parses/pdf_parses_0.jsonl
       |-- sample.jsonl
|-- RELEASE_NOTES           # a few notes about this particular release
|-- LICENSE                 # license for usage of S2ORC

Data size

We strongly recommend users to first download the sample/ (5.9MB) which before committing to downloading the full/ collection (52GB for metadata/ and 160GB for pdf_parses/).

Also make sure there is enough space on your machine. When unzipped the dataset grows to 191GB for metadata/ and 643GB for pdf_parses/.

gzip

Each JSONlines file in full/ is zipped up. To unzip all files in the directory: cd <dirname> and gunzip *.

metadata schema

We recommend everyone work with metadata/ as the starting point. This is a JSONlines file (one line per paper) with the following keys:

Identifier fields

• paper_id: a str-valued field that is a unique identifier for each S2ORC paper.

• arxiv_id: a str-valued field for papers on arXiv.org.

• acl_id: a str-valued field for papers on the ACL Anthology.

• pmc_id: a str-valued field for papers on PubMed Central.

• pubmed_id: a str-valued field for papers on PubMed, which includes MEDLINE. Also known as pmid on PubMed.

• mag_id: a str-valued field for papers on Microsoft Academic.

• doi: a str-valued field for the DOI.

Notably:

• Resolved citation links are represented by the cited paper's paper_id.

• The paper_id resolves to a Semantic Scholar paper page, which can be verified using the s2_url field.

• We don't always have a value for every identifier field. When missing, they take null value.

Metadata fields

• title: a str-valued field for the paper title. Every S2ORC paper must have one, though the source can be from publishers or parsed from PDFs. We prioritize publisher-provided values over parsed values.

• authors: a List[Dict]-valued field for the paper authors. Authors are listed in order. Each dictionary has the keys first, middle, last, and suffix for the author name, which are all str-valued with exception of middle, which is a List[str]-valued field. Every S2ORC paper must have at least one author.

• venue and journal: str-valued fields for the published venue/journal. Please note that there is not often agreement as to what constitutes a "venue" versus a "journal". Consolidating these fields is being considered for future releases.

• year: an int-valued field for the published year. If a paper is preprinted in 2019 but published in 2020, we try to ensure the venue/journal and year fields agree & prefer non-preprint published info. We know this decision prohibits certain types of analysis like comparing preprint & published versions of a paper. We're looking into it for future releases.

• abstract: a str-valued field for the abstract. These are provided directly from gold sources (not parsed from PDFs). We preserve newline breaks in structured abstracts, which are common in medical papers, by denoting breaks with ':::'.

• inbound_citations: a List[str]-valued field containing paper_id of other S2ORC papers that cite the current paper. Currently derived from PDF-parsed bibliographies, but may have gold sources in the future.

• outbound_citations: a List[str]-valued field containing paper_id of other S2ORC papers that the current paper cites. Same note as above.

• has_inbound_citations: a bool-valued field that is true if inbound_citations has at least one entry, and false otherwise.

• has_outbound_citations a bool-valued field that is true if outbound_citations has at least one entry, and false otherwise.

We don't always have a value for every metadata field. When missing, str fields take null value, while List fields are empty lists.

PDF parse-related metadata fields

• has_pdf_parse: a bool-valued field that is true if this paper has a corresponding entry in pdf_parses/, which means we had processed that paper's PDF(s) at some point. The field is false otherwise.

• has_pdf_parsed_abstract: a bool-valued field that is true if the paper's PDF parse contains a parsed abstract, and false otherwise.

• has_pdf_parsed_body_text: a bool-valued field that is true if the paper's PDF parse contains parsed body text, and false otherwise.

• has_pdf_parsed_bib_entries: a bool-valued field that is true if the paper's PDF parse contains parsed bibliography entries, and false otherwise.

• has_pdf_parsed_ref_entries: a bool-valued field that is true if the paper's PDF parse contains parsed reference entries (e.g. tables, figures), and false otherwise.

Please note:

• If has_pdf_parse = false, the other four fields will not be present in the JSON (trivially false).

• If has_pdf_parse = true but has_pdf_parsed_abstract, has_pdf_parsed_body_text, or has_pdf_parsed_ref_entries are false, this can be because:

  • Our PDF parser failed to extract that element
  • Our PDF parser succeeded but that paper simply did not have that element (e.g. papers without abstracts)
  • Our PDF parser succeeded but that element was removed because the paper is not identified as open-access.

pdf_parses schema

We view pdf_parses/ as supplementary to the metadata/ entries. PDF parses are also represented as JSONlines file (one line per paper) with the following keys:

• paper_id: a str-valued field which is the same S2ORC paper ID in metadata/

• _pdf_hash: a str-valued field. Internal usage only. We use this for debugging.

• abstract and body_text are List[Dict]-valued fields representing parsed text from the PDF. Each Dict corresponds to a paragraph. List preserves their original ordering.

• bib_entries and ref_entries are Dict-valued fields representing extracted entities that can be referenced (inline) within the text.

example 1

One example paragraph in abstract or body_text might look like:

{
    "section": "Introduction",
    "text": "Dogs are happier cats [13, 15]. See Figure 3 for a diagram.",
    "cite_spans": [
        {"start": 22, "end": 25, "text": "[13", "ref_id": "BIBREF11"},
        {"start": 27, "end": 30, "text": "15]", "ref_id": "BIBREF30"},
        ...
    ],
    "ref_spans": [
        {"start": 36, "end": 44, "text": "Figure 3", "ref_id": "FIGREF2"},
    ]
}

and example bib_entries and ref_entries might look like:

{
    ...,
    "BIBREF11": {
        "title": "Do dogs dream of electric humans?",
        "authors": [
            {"first": "Lucy", "middle": ["Lu"], "last": "Wang", "suffix": ""}, 
            {"first": "Mark", "middle": [], "last": "Neumann", "suffix": "V"}
        ],
        "year": "", 
        "venue": "barXiv",
        "link": null
    },
    ...
}

{
    "TABREF4": {
        "text": "Table 5. Clearly, we achieve SOTA here or something.",
        "type": "table"
    }
    ...,
    "FIGREF2": {
        "text": "Figure 3. This is the caption of a pretty figure.",
        "type": "figure"
    },
    ...
}

Notice:

• Inline spans are represented by character start and end indices into the paragraph text
• spans resolve to BIBREF, TABREF or FIGREF entries.
• BIBREF are IDs of bibliographic elements of bib_entries. Bib entries may be missing fields (e.g. year). They can be linked to S2ORC papers, as specified by link, but we also preserve any unlinked entries by setting link to null.
• FIGREF and TABREF are IDs of figure and table elements of ref_entries. Ref entries contain the caption text of the corresponding object, and also indicate the type of object.

example 2

You may see empty pdf_parses/ JSONs that look like:

{
    "paper_id": "...", 
    "_pdf_hash": "...", 
    "abstract": [], 
    "body_text": [], 
    "bib_entries": {}, 
    "ref_entries": {}
}

We keep these around for our internal usage, but the way to interpret these is that there is no usable PDF parse here, despite the corresponding metadata/ entry still displaying has_pdf_parse = true.

These exist when (i) bib_entries does not successfully parse and (ii) the paper is not open-access, so we had to remove abstract, body_text, and ref_entries.

Example Usage

To start with something simple, let's count (outbound) citation edges:

import os
import json

num_edges = 0
METADATA_DIR = 'full/metadata/'
for metadata_file in os.listdir(METADATA_DIR):
    with open(os.path.join(METADATA_DIR, metadata_file)) as f_meta:
        for line in f_meta:
            metadata_dict = json.loads(line)
            num_edges += len(metadata_dict['outbound_citations'])
print(num_edges)
> 467588220

The way we expect most people to use S2ORC full text is to loop through the metadata/ entries, select papers of interest, and retrieve the corresponding pdf_parses/ entries when desired. For example:

import json

# feel free to wrap this into a larger loop for batches 0~99
BATCH_ID = 0

# create a lookup for the pdf parse based on paper ID
paper_id_to_pdf_parse = {}
with open(f'full/pdf_parses/pdf_parses_{BATCH_ID}.jsonl') as f_pdf:
    for line in f_pdf:
        pdf_parse_dict = json.loads(line)
        paper_id_to_pdf_parse[pdf_parse_dict['paper_id']] = pdf_parse_dict

# filter papers using metadata values
citation_contexts = []
with open(f'full/metadata/metadata_{BATCH_ID}.jsonl') as f_meta:
    for line in f_meta:
        metadata_dict = json.loads(line)
        paper_id = metadata_dict['paper_id']
        print(f"Currently viewing S2ORC paper: {paper_id}")
        
        # suppose we only care about ACL anthology papers
        if not metadata_dict['acl_id']:
            continue
    
        # and we want only papers with resolved outbound citations
        if not metadata_dict['has_outbound_citations']:
            continue
        
        # get citation context (paragraphs)!
        if paper_id in paper_id_to_pdf_parse:
            # (1) get the full pdf parse from the previously computed lookup dict
            pdf_parse = paper_id_to_pdf_parse[paper_id]
            
            # (2) pull out fields we need from the pdf parse, including bibliography & text
            bib_entries = pdf_parse['bib_entries']
            paragraphs = pdf_parse['abstract'] + pdf_parse['body_text']
            
            # (3) loop over paragraphs, grabbing citation contexts
            for paragraph in paragraphs:
                
                # (4) loop over each inline citation in this paragraph
                for cite_span in paragraph['cite_spans']:
                    
                    # (5) each inline citation can be resolved to a bib entry
                    cited_bib_entry = bib_entries[cite_span['ref_id']]
                    
                    # (6) that bib entry *may* be linked to a S2ORC paper.  if so, grab paragraph
                    linked_paper_id = cited_bib_entry['link']
                    if linked_paper_id:
                        citation_contexts.append({
                            'citing_paper_id': paper_id,
                            'cited_paper_id': linked_paper_id,
                            'context': paragraph['text'],
                            'citation_mention_start': cite_span['start'],
                            'citation_mention_end': cite_span['end'],
                        })

Contact us

The best way to contact us is through email. Don't hesitate to reach out about anything; we've helped a lot of people get started with the dataset, which can be a bit daunting given its size.

Email: {kylel, lucyw}@allenai.org

Twitter @kylelostat, @lucyluwang

IRC: #s2orc at irc.oftc.net

Give us Feedback: Totally optional, but we'd love to hear how you're using this dataset & any feedback for improving it. Send us an email or use this Google Form.

Mailing list: Join our Google Groups! We email updates & releases about S2ORC here.

Report issues: Use GitHub Issues to report bugs or issues! We'll try to fix it for the next release.

FAQ

How is this related to the other Semantic Scholar Open Research Corpus?

S2ORC and the older Semantic Scholar Open Research Corpus should be viewed as separate datasets for now, though we're hoping to combine them into a single service in the long-term. Until then, both are available because there are many existing projects that still rely on one or the other. Differences between the two:

Definitions & Scope

• S2ORC uses a different way of selecting canonical metadata fields & prioritizing the "version" of the paper to use as canonical (e.g. camera-ready vs preprint).
• S2ORC has a stricter notion "paper" criteria and does not include slides, dissertations, books, etc.. It also currently does not include non-English papers. Therefore, S2ORC will have fewer documents in its collection than the other dataset.

Citation graph

• S2ORC uses a different bibliography parser & linker that's higher precision but lower coverage. Consolidating the two is on our TODO list.

Metadata

• S2ORC does not provide resolved Author IDs, just surface forms.
• S2ORC provides additional IDs not provided in the other dataset, like arXiv IDs and PMC IDs.

PDF Parses

• S2ORC provides machine-readable full text for open-access papers, while the other does not provide any full text.

Release & Access

• S2ORC is released under CC BY-NC 2.0, which is only for non-commercial use, while the other is released under an ODC-BY license, which allows commercial use.
• S2ORC is updated at an ad hoc basis now, while the other is automated with monthly dumps.

If you're unsure what to use, please email us and we'd be happy to discuss your project with you.

License

S2ORC is released under the CC BY-NC 2.0. By using S2ORC, you are agreeing to its usage terms.

Citation

If using this dataset, please cite:

@inproceedings{lo-wang-2020-s2orc,
    title = "{S}2{ORC}: The Semantic Scholar Open Research Corpus",
    author = "Lo, Kyle  and Wang, Lucy Lu  and Neumann, Mark  and Kinney, Rodney  and Weld, Daniel",
    booktitle = "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
    month = jul,
    year = "2020",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/2020.acl-main.447",
    doi = "10.18653/v1/2020.acl-main.447",
    pages = "4969--4983"
}