In-Depth Survey of Open Paper Knowledge Graphs on the Internet

In-Depth Survey of Open Paper Knowledge Graphs on the Internet

*A research report for selecting data resources for a paper-writing assistance website. Survey date: 2026-07-10; method: four parallel deep-search agents with cross-verification of key facts; scope: 22 Chinese and English open resources.*

Executive Summary

• Primary backbone: OpenAlex (CC0, ~477M works, REST API + quarterly snapshots, actively maintained) — best overall coverage and licensing for a citation/literature index.
• DOI and citation backbone: Crossref (CC0, 176M DOIs) + OpenCitations/COCI (CC0, 2.2B citation edges).
• Full text & semantics: S2ORC, arXiv, PubMed/PMC for full-text mining; Semantic Scholar (S2AG) provides TLDRs and citation recommendations, but bulk data is non-commercial (ODC-BY).
• Methods/tasks/datasets: Papers with Code was shut down by Meta in July 2025; data frozen in a GitHub archive; rebuilt in 2026 by paperswithcode.co with AI agents.
• Chinese gap is significant: only NCPSSD (13.5M Chinese social-science full texts) and AMiner/OAG are genuinely usable; CNKI, Wanfang, VIP, and CSCD are all paywalled or closed.
• Biggest risk: scarce, closed-licensed Chinese paper data plus full-text copyright and citation lag — rely on compliant international sources as the core, with Chinese data as limited enhancement only.

Key Findings by Resource (English)

Comprehensive scholarly graphs

• OpenAlex (OurResearch, nonprofit): ~477M works, 116M authors, 100K institutions, 65K concepts, 2B+ citations. Entities: works/authors/sources/institutions/topics/funders/geo. Access: REST API + quarterly snapshots, no key needed (mailto polite pool). CC0, commercially usable, highly active. Rating: ★★★★★
• Semantic Scholar (S2AG) (Allen Institute for AI): 214M+ papers, 79M authors, 2.49B citations, SPECTER2 embeddings. API is CC0; bulk datasets ODC-BY, non-commercial. Rating: ★★★★☆
• Microsoft Academic Graph (MAG): discontinued 2021-12-31; succeeded by OpenAlex. Do not use for new projects.
• Open Academic Graph (OAG) (Tsinghua AMiner + Microsoft Research): v3.1 (2024-02) with ~700M entities, 2B relations. Bulk download, ODC-BY, research-oriented. Rating: ★★★☆☆
• DBLP (Schloss Dagstuhl): ~8.1M records, 3.9M authors, CS-focused. Monthly XML dumps + SPARQL, CC0. No citation network. Rating: ★★★☆☆
• arXiv (Cornell): 2.4M+ e-prints. OAI-PMH + Atom API; metadata CC0; preprints, not peer-reviewed. Rating: ★★★★☆
• PubMed/PMC (NCBI/NLM): 39M+ citations; 11M+ PMC full texts. E-utilities + FTP. Public-domain metadata. Rating: ★★★★☆
• Crossref: 176M+ DOI records, annual public data files (197GB). CC0. Rating: ★★★★★

Citation and supplementary graphs

• OpenCitations/COCI: 2.2B+ citations (as of July 2025) from Crossref, NIH-OCC, DataCite. REST + SPARQL + CSV/N-Triples, no key. CC0. Excellent for citation network analysis. Rating: ★★★★☆
• WikiCite/Wikidata: ~41M scholarly-article items; dedicated SPARQL endpoint since May 2025 (~8B triples). Properties: cites work (P2860), author (P50), published in (P1433), DOI (P356). CC0. Rating: ★★★☆☆
• Papers with Code: shut down by Meta 2025-07-24/25; archived on GitHub; rebuilt in 2026 at paperswithcode.co with AI agents, restoring SOTA leaderboards. Rating: ★★★★☆
• Connected Papers: commercial, built on Semantic Scholar; no open API, commercial use restricted. Rating: ★★☆☆☆
• S2ORC (Allen AI): 81M paper nodes, 8.1M structured OA full texts, 467M+ citation edges. CC0 (restricted). Best for full-text mining. Rating: ★★★☆☆
• CiteSeerX (Penn State): 10M+ documents, 100M+ citations (CS). REST + OAI-PMH + S3 snapshots; aging. Rating: ★★☆☆☆
• ORKG (TIB, Germany): structured problem–method–result descriptions. REST + RDF; CC BY-SA. Rating: ★★★☆☆
• Closed counterexamples: Google Scholar (no API, automation banned), ResearchGate (no open API), CNKI (commercial, no public API).

Chinese open resources

• AMiner/OAG Chinese: OAG dataset genuinely downloadable (ODC-BY) but low Chinese-paper share; AMiner has 331M papers, 135M scholars with strong Chinese author/institution coverage. Rating: ★★★★☆
• OpenKG.cn: open community KG hub (337 datasets + 71 tools); academic subgraphs (SciKG, CN-DBpedia, etc.) are mostly English arXiv or encyclopedia-based — no dedicated Chinese-paper KG; useful as schema reference. Rating: ★★☆☆☆
• NCPSSD (National Center for Philosophy and Social Sciences Documentation): genuinely open access — free PDF reading/download after registration; 25M records total, 13.5M+ Chinese social-science papers. No open API; compliance must be assessed. Rating: ★★★★☆
• NSTL: free search but full text via document delivery with daily quotas; not suitable as an open data foundation.
• CSCD: institutional subscription only (WoS version); not open.
• Wanfang / VIP / CNKI: all closed — subscription/paywall, no public API, bulk scraping prohibited. Rating: ✕
• CAS GoOA: OA journal discovery (1,700+ journals) downloadable; its scholarly KG (400M+ entities, 6.05B+ relations) is not open.
• Fudan CN-DBpedia: 9M entities, 67M triples; dump downloadable under research license; encyclopedia-based, not a paper KG.

> Chinese takeaway: Truly open: OAG, OpenKG subgraphs, NCPSSD, GoOA, CN-DBpedia. Search-only/closed: CSCD, Wanfang/VIP/CNKI, NSTL delivery, CAS internal KG. A truly open Chinese scholarly paper knowledge graph remains a blank spot. Recommended combination: NCPSSD full text + OAG citations + OpenKG schema.

Comparison Matrix (abridged)

| Resource | License | Chinese coverage | Active | Writing fitness | |---|---|---|---|---| | OpenAlex | CC0 | Medium | High | ★★★★★ | | Semantic Scholar (S2AG) | CC-BY-NC (bulk) | Weak | High | ★★★★☆ | | MAG | Discontinued | Weak | Low | ✕ | | OAG | ODC-BY | Medium | Medium | ★★★☆☆ | | DBLP | CC0 | Weak | High | ★★★☆☆ | | arXiv | CC0 (metadata) | Weak | High | ★★★★☆ | | PubMed/PMC | Public domain | Weak | High | ★★★★☆ | | Crossref | CC0 | Medium | High | ★★★★★ | | OpenCitations/COCI | CC0 | Weak | High | ★★★★☆ | | WikiCite/Wikidata | CC0 | Medium | High | ★★★☆☆ | | Papers with Code | MIT/CC-BY | Weak | Medium | ★★★☆☆ | | Connected Papers | Commercial-restricted | Weak | High | ★★☆☆☆ | | S2ORC | CC0 (limited) | Weak | Low | ★★★☆☆ | | ORKG | CC BY-SA | Weak | Medium | ★★★☆☆ | | AMiner/OAG | Research license | Strong | Medium | ★★★★☆ | | OpenKG.cn | Mixed | Strong | Medium | ★★☆☆☆ | | NCPSSD | Domestic authorization | Strong | Medium | ★★★★☆ | | CSCD | Commercial license | Strong | Medium | ★★★☆☆ | | Wanfang/VIP/CNKI | Paywall | Strong | High | ✕ | | CAS GoOA | OA (discovery) | Strong | Medium | ★★☆☆☆ |

Recommended Architecture

1. Index & citation core: OpenAlex (works/authors/institutions/topics) + Crossref (DOI normalization) + OpenCitations (open citation edges). 2. Semantic/full-text layer: S2ORC + arXiv + PubMed/PMC for full-text mining; Semantic Scholar API for TLDRs/embeddings (non-commercial limits apply). 3. Chinese enhancement: NCPSSD full text + OAG author/institution data + OpenKG schemas. 4. Task/method links: paperswithcode.co archives, cautiously, given instability.

Risks

• Chinese open scholarly data is scarce and licensing-closed; avoid making Chinese sources the core dependency.
• Full-text copyright and citation-index lag affect timeliness.
• Third-party service shutdowns (MAG, Papers with Code) show vendor risk — prefer CC0, actively maintained infrastructure like OpenAlex and OpenCitations.