Research Knowledge Graphs (RKGs)

• Research Knowledge Graphs (RKGs) are semantically rich, graph-based data structures that encode research artifacts and their typed relationships using persistent identifiers, ontologies, and FAIR principles.
• They employ diverse construction methodologies—including manual curation, rule-based extraction, deep learning pipelines, and hybrid approaches—to ensure scalable and accurate integration of scholarly metadata.
• RKGs support advanced query answering and analytics via SPARQL endpoints, similarity-based recommendations, and neuro-symbolic interfaces, improving discoverability and reproducibility in research.

Research Knowledge Graphs (RKGs) are semantically rich, graph-based data structures that encode research entities and their typed relationships in a machine-actionable form. Across the surveyed literature, they are defined as directed, labeled graphs or multigraphs tailored to research artifacts such as publications, datasets, software, models, methods, persons, organizations, themes, and results. Their primary objectives are transparency, reproducibility, discoverability, and the realization of FAIR principles through standard ontologies, shared vocabularies, persistent identifiers, and queryable interlinking of scholarly resources (Zloch et al., 8 Jun 2025, Silva et al., 4 Aug 2025, Jaradeh et al., 2019).

1. Formal definitions and representational scope

A general formalization models an RKG as

$$\mathrm{RKG}=(V,E,\mathcal T,\mathcal R,L,\mathrm{PID}),$$

where $V$ is the set of nodes, $\mathcal R$ is the set of relation types, $E \subseteq V \times \mathcal R \times V$ is the set of directed, typed edges, $\mathcal T$ is a hierarchy of node-types, $L$ assigns human-readable labels, and $\mathrm{PID}$ assigns each node a globally unique, persistent identifier such as a DOI, ORCID, or ROR (Zloch et al., 8 Jun 2025). This formulation emphasizes that research artifacts are not merely linked; they are typed, labeled, and persistently identified.

A complementary formulation used in NFDI4DataScience represents an RKG as

$G=(V,E,\tau),$

where $V$ is the set of nodes, $E \subseteq V \times V$ is the set of directed edges, and $V$0 maps nodes and edges to ontology classes and properties (Silva et al., 4 Aug 2025). In this model, a dataset node may be typed as nfdids:Dataset, a model node as nfdids:Model, and an edge between them as nfdids:usesDataset. The significance given in the source is explicit: such a formal model supports query answering, reasoning, and the seamless integration of heterogeneous metadata sources.

Other formulations are application-specific. In the literature-review setting, EmpiRE-Compass defines an RKG as a directed, labeled multi-graph $V$1 in which $V$2 contains research entities such as papers, themes, methods, and results, and $V$3 contains labeled triples such as addressesTheme, usesMethod, yieldsResult, and citesPaper (Karras et al., 25 Feb 2026). In the context of predicate quality management, the ORKG-oriented framework formalizes an RKG as $V$4 with $V$5 as the set of predicates and $V$6 as the set of triples (John et al., 3 Jul 2026). Taken together, these variants show that RKGs are not a single schema but a family of formally compatible graph models specialized for research artifacts, provenance, and curation workflows.

2. Ontological foundations, identifiers, and modeling principles

The core principles recurring across RKG work are the use of non-proprietary, Web-resolvable Persistent Identifiers for all entities, reuse of established vocabularies and ontologies, interlinking across graphs via shared URIs, adherence to FAIR principles, and clear versioning and provenance tracking (Zloch et al., 8 Jun 2025). These principles distinguish RKGs from ad hoc metadata stores: identifiers and vocabulary reuse are treated as constitutive elements, not optional enhancements.

NFDI4DataScience structures its RKG ecosystem in four logical layers: Community Benchmarking; Tooling for Information Extraction (IE) & FAIR Assessment; Vocabularies, Schemas & Ontologies; and the RKGs themselves (Silva et al., 4 Aug 2025). Within this architecture, the NFDICore Ontology is a mid-level, BFO-compliant ontology modeling generic NFDI resources, with key classes such as nfdicore:Resource, nfdicore:Person, nfdicore:Organization, nfdicore:Dataset, and nfdicore:Software, and properties such as nfdicore:hasAuthor and nfdicore:hasLicense. The NFDI4DS Ontology extends NFDICore with domain-specific classes including nfdids:MLModel and nfdids:EvaluationMetric, and properties such as nfdids:isDerivedFrom and nfdids:hasTemporalVersion. The metadata standards explicitly listed are DDI Alliance, DataCite, schema.org, ORCID, and ROR.

The Open Research Knowledge Graph adopts a lighter core ontology centered on four primary concepts: ResearchContribution, ResearchProblem, ResearchMethod, and ResearchResult (Jaradeh et al., 2019). Each ResearchContribution links via addresses to one or more ResearchProblem nodes, via uses or employs to ResearchMethod nodes, and via yields to ResearchResult nodes. The same source states that ORKG does not force users into a monolithic, all-encompassing ontology; at the leaves of the graph users are free to adopt any external vocabulary, and imported RDF triples can be mapped into the labeled-property graph and exported to RDF/OWL for SPARQL querying and reasoning. This combination of lightweight core concepts and external vocabulary reuse suggests a design pattern in which interoperability is achieved through alignment rather than schema closure.

3. Construction methodologies and infrastructural patterns

The surveyed literature distinguishes four main construction methodologies: manual curation, rule-based ontology-driven extraction, deep learning / neural-symbolic pipelines, and hybrid approaches (Zloch et al., 8 Jun 2025). Manual curation offers very high precision but low scalability. Rule-based extraction maps text to RDF through functions for metadata mapping, sentence segmentation, ontology-based Named Entity Recognition, dependency-rule relation extraction, and URI disambiguation. Deep learning / neural-symbolic pipelines use transformers such as BERT or SciBERT together with symbolic post-processing, classification, linking to external PIDs, and sampled validation. Hybrid approaches combine manual or community correction with automatic extraction for scalability and quality.

ORKG is an explicit implementation of multi-modal scholarly knowledge acquisition, combining crowdsourced and automated techniques in a single unified workflow (Jaradeh et al., 2019). Its front-end wizard guides authors through four steps: select the paper by DOI, enter the ResearchProblem, specify the ResearchMethod, and record the ResearchResult. In a user evaluation with twelve DILS2018 authors, the average time to produce a structured contribution was 17 min, with a minimum of 13 min and a maximum of 22 min, yielding on average 53 triples per contribution. Automated extraction in the same infrastructure uses SciBERT embeddings in a Bi-LSTM + CRF architecture trained on the Elsevier OA-STM corpus to recognize PROCESS, METHOD, MATERIAL, and DATA, followed by entity linking with Falcon, Spotlight, TAGME, EARL, TextRazor, and MeaningCloud. On the STM test corpus, Falcon achieved the highest coverage $V$7.

NFDI4DataScience describes a broader infrastructural stack in which IE pipelines harvest metadata and content from repositories including Hugging Face, arXiv, GitHub, and Zenodo; NLP methods extract entity mentions from PDFs and HTML; Entity Linking aligns mentions to canonical identifiers such as DOI, Model ID, and ORCID; shared tasks such as SOMD-2024/2025 benchmark software-mention detection; and FAIR assessment pipelines, including LLM-based evaluators, score artifacts on metrics like metadata completeness, PID usage, and licensing clarity (Silva et al., 4 Aug 2025).

EmpiRE-Compass exemplifies a neuro-symbolic pattern specialized for literature reviews (Karras et al., 25 Feb 2026). Initial LR data in CSVs or spreadsheets is transformed into RDF triples via transformation scripts using Python and RDFLib, then stored in ORKG via its HTTP-RDF API. The system separates a Symbolic Layer, based on SPARQL and ORKG, from a Neural Layer based on LLMs. For curated competency questions, the system executes one of 16 pre-defined SPARQL templates and renders the result through D3.js, chart libraries, and React components. For custom competency questions, natural-language input is translated into draft SPARQL, which remains editable by the researcher. The source explicitly recommends allowing manual query inspection and logging and exporting every stage—prompts, queries, raw results, visualizations, and narratives—to ensure full reproducibility.

4. Representative implementations and domain coverage

The RKG landscape described in the literature is heterogeneous with respect to scale, scope, data sources, vocabularies, and reliability (Zloch et al., 8 Jun 2025). Examples span cross-domain scholarly metadata, domain-specific scientific corpora, manually curated ground truth, primary research data, automatically extracted relations, and community-curated representations of research contributions.

In NFDI4DataScience, specialized RKGs are grouped by focus (Silva et al., 4 Aug 2025). Scholarly Resource Metadata includes the dblp KG with 510 M+ RDF triples, 3.7 M authors, 7.8 M publications, and a public SPARQL service; the GESIS KG with metadata on social-science datasets, publications, and survey variables; gesisDataSearchKG; a KG of AI & DS Methods in planning; and a Hugging Face KG in development. Quality-Controlled Ground Truth includes SoMeSci KG with 3,756 manually annotated software mentions in 1,367 PMC articles, and GSAP KG with curated annotations of ML models and datasets in 100 CS papers. Primary Research Data includes TweetsKB with 3.1 B tweets from 2013–2023, TweetsCOV19 with 41 M COVID-19 tweets, and ClaimsKG with 74 k fact-check claims linked to DBpedia concepts. NLP-Derived Resource Relations include SoftwareKG with software mentions from 51 k social-science articles. Community-Curated Scholarly Expressions include ORKG.

The STM-KG used for citation recommendation is a paper–concept graph over ten STM domains built from 55 485 abstracts and 15 395 intra-KG citation edges, of which 2 200 are cross-domain (Brack et al., 2021). It has paper nodes, concept nodes, mentions edges from papers to concepts, and citation links between papers. The graph is instantiated in two variants: a cross-domain KG with approximately 994 572 total concepts and an in-domain KG with a summed concept vocabulary of approximately 1 178 480. For evaluation, only papers with at least 4 citations within the KG are used, yielding 720 query papers and 4 069 valid citation links.

For topic detection and prediction, the AIDA knowledge graph integrates Microsoft Academic Graph, GRID, DBpedia, and optionally the Dimensions API, and scales to approximately 21 M publications and approximately 8 M patents (Salatino et al., 2021). Its schema includes cso:Topic, cso:ResearchPaper, cso:Patent, aida:IndustrialSector, and aida:AffiliationType, together with topic, citation, affiliation-type, and industrial-sector relations. This range of implementations indicates that the term RKG denotes a scholarly-knowledge modeling paradigm rather than a single platform or community.

5. Querying, analytics, and downstream uses

A central function of RKGs is query answering over interconnected scholarly entities. NFDI4DataScience lists SPARQL Endpoints & SHMARQL Interfaces, visualization dashboards, ingestion pipelines, and FAIR Assessment Services as core tools and services (Silva et al., 4 Aug 2025). The applications enumerated in the broader survey include state-of-the-art discovery, reproducibility support, citation and credit tracking, scholarly search and recommendation, and LLM augmentation (Zloch et al., 8 Jun 2025).

ORKG operationalizes these capabilities through state-of-the-art comparison and similarity-based recommendation (Jaradeh et al., 2019). For comparison across contributions, it computes predicate-embedding cosine similarities $V$8, builds a binary mask over predicate occurrence, and aligns semantically similar predicates with an empirical threshold $V$9. The reported runtime is approximately 20 ms even for 8 contributions, versus a baseline of approximately 14421 s for 8 contributions. For retrieval, each contribution subgraph is linearized as a document of subject–predicate–object label concatenations, indexed with TF/IDF, and queried for top-$\mathcal R$0 contributions. The same infrastructure also supports SPARQL querying and frontend visualizations as tables, CSV exports, LaTeX tables, or network graphs.

In citation recommendation, the STM-KG contributes an additional signal beyond paper text and citation structure (Brack et al., 2021). For each paper $\mathcal R$1, a sparse concept vector

$\mathcal R$2

is concatenated with a dense document embedding $\mathcal R$3 to obtain

$\mathcal R$4

Relevance for a query paper $\mathcal R$5 and a candidate $\mathcal R$6 is then computed by cosine similarity,

$\mathcal R$7

The strongest baseline, SPECTER, obtains MAP@50 of 19.8, while SPECTER + concepts reaches 20.6, a gain of +0.8. Concept vectors alone yield MAP@50 around 9 %, far above random at 0 %.

RKGs also support topic analysis and forecasting. The topic-oriented framework built on CSO and AIDA uses syntactic and semantic topic assignment, clique-percolation and Advanced CPM on topic co-occurrence graphs, bibliometric analyses over citation and co-citation patterns, and forecasting models including Logistic Regression, Random Forest, AdaBoost, CNN, and LSTM (Salatino et al., 2021). On a gold standard of 9 776 samples, with 10-fold cross-validation, the best result is obtained by LSTM with all four signals—papers from Academia, Industry, and patents from Academia, Industry—achieving Precision = 83.2 %, Recall = 86.1 %, and $\mathcal R$8.

EmpiRE-Compass extends querying with competency-question workflows and neuro-symbolic synthesis (Karras et al., 25 Feb 2026). It offers 16 pre-defined competency questions, each paired with a SPARQL template, and a custom workflow in which a natural-language question is converted into draft SPARQL, executed, visualized, and narratively summarized. All SPARQL queries, LLM prompts, retrieved data, visualizations, interpretations, explanations, and user-interaction histories are exportable as a JSON bundle, and import functionality allows resuming an analysis session. This makes the query itself a reusable research artifact.

6. Data quality, predicate redundancy, and curator-centered governance

Weakly enforced schemas make RKGs prone to inconsistencies, particularly in how predicates are defined and used (John et al., 3 Jul 2026). The predicate-redundancy framework formalizes duplicate predicates as distinct identifiers $\mathcal R$9 expressing the same or highly similar semantic relation. Given a similarity function

$E \subseteq V \times \mathcal R \times V$0

and a threshold $E \subseteq V \times \mathcal R \times V$1, $E \subseteq V \times \mathcal R \times V$2 and $E \subseteq V \times \mathcal R \times V$3 are duplicates if

$E \subseteq V \times \mathcal R \times V$4

Under a closed-world assumption for the observed schema, the goal is to identify all such pairs and consolidate them to improve the conciseness of $E \subseteq V \times \mathcal R \times V$5.

The duplicate-detection phase represents each predicate by concatenating its rDFS label and description into a text sequence $E \subseteq V \times \mathcal R \times V$6, computing embeddings $E \subseteq V \times \mathcal R \times V$7 with the pretrained SentenceTransformers model all-MiniLM-L6-v2, where $E \subseteq V \times \mathcal R \times V$8, and measuring cosine similarity

$E \subseteq V \times \mathcal R \times V$9

A $\mathcal T$0-nearest-neighbor graph is built on predicates with $\mathcal T$1 by connecting $\mathcal T$2 to $\mathcal T$3 if $\mathcal T$4, and clusters are extracted via the Louvain community detection algorithm to maximize modularity. In SciKGDash, curators inspect each cluster through labels, descriptions, usage counts, provenance, and an embedding-similarity graph. Low-cost heuristic flags suggest delete for predicates with zero usage, merge for predicates with identical normalized label, and review otherwise. Curators then select one of four actions per cluster: delete, distinguish, soft-merge, or hard-merge.

The resolution semantics are explicitly formalized. A hard-merge replaces all triples using duplicates by a chosen canonical predicate $\mathcal T$5:

$\mathcal T$6

A soft-merge adds equivalence axioms without modifying $\mathcal T$7, and deletion of an unused predicate removes it from $\mathcal T$8 and all incident triples from $\mathcal T$9. Preventive mechanisms include real-time similarity-based autocomplete with threshold $L$0, warning dialogs for equal normalized labels or high similarity, metadata-enrichment guidelines requiring concise, unique labels and descriptions, a centralized registry of canonical predicates such as hasMethod and hasData, enforced naming conventions of lowerCamelCase with no underscores and one verb phrase per predicate, and validation hooks in all ingestion paths.

Evaluation on ORKG makes the quality issue concrete. The predicate set has $L$1, and average usage per predicate is 4,000. Clustering produced 3,837 clusters: 2,769 singletons and 1,068 multi-predicate clusters. Heuristic suggestions marked 852 clusters containing 3,185 predicates, approximately 23.7 %, for merge; 338 clusters containing 426 predicates, approximately 3.2 %, for deletion; and 162 clusters containing 7,361 predicates, approximately 54.7 %, for manual review. In total, 4,033 predicates, approximately 30.0 % of $L$2, were flagged as potentially removable. After applying hard-merges and deletions on sandbox data, the predicate count was reduced from 13,457 to approximately 9,424, the coherence score improved from 72 % to over 90 %, and query answering benchmarks on common questions showed a 15 % average speedup and a 10 % increase in recall. The same study reports recurring modeling patterns that lead to predicate redundancy, user-induced duplication, inconsistent identifier usage, and a lack of standardization in predicate naming and usage; duplicate predicates arise from user behavior and interface design. This places data quality management at the center of RKG maintenance rather than at the periphery.

7. Misconceptions, open challenges, and future directions

A recurrent misconception is that an RKG is simply a general-purpose knowledge graph populated with research papers. The literature-review formulation explicitly distinguishes RKGs from standard knowledge graphs such as DBpedia and Wikidata by their explicit modeling of competency questions driving the review, provenance links to extraction workflows, tight integration with secondary-study processes, and versioning and continuous update support for living reviews (Karras et al., 25 Feb 2026). More generally, the RKG survey emphasizes persistent identifiers, provenance, and research-artifact relations as defining features rather than optional annotations (Zloch et al., 8 Jun 2025).

Open challenges are consistent across the corpus. The broad survey lists scalability, heterogeneity, quality control, versioning and temporal reasoning, and incentives and community adoption (Zloch et al., 8 Jun 2025). ORKG adds the tension between granularity and cost, asking how finely a research contribution can be sliced before curation overhead becomes prohibitive, and notes future needs such as full SPARQL 1.1/OWL2 support, nanopublication export, integration with Papers With Code and Semantic Scholar dumps, lifecycle integration with Jupyter-based data analysis, faceted browsing over predicate hierarchies, and probabilistic querying over uncertain triples (Jaradeh et al., 2019). NFDI4DataScience identifies future directions in federated querying across NFDI4DS RKGs and other NFDI consortia, deeper provenance modeling through fine-grained versioning and isDerivedFrom chains, harmonization of metadata from GitHub, Zenodo, Hugging Face, and arXiv into a unified RKG backbone, enhanced reasoning services to detect reproducibility gaps, and community engagement via ontology hackathons and additional shared tasks (Silva et al., 4 Aug 2025).

Neuro-symbolic systems introduce further challenges. EmpiRE-Compass reports that ensuring LLM-generated SPARQL correctness and schema compliance requires iterative refinement and guardrails, that trustworthiness of LLM-generated interpretations must be validated by domain experts before release, and that balancing real-time performance with API cost constraints demands caching strategies and request quotas (Karras et al., 25 Feb 2026). These issues are not external to RKGs; they arise precisely because RKGs are increasingly used as the symbolic substrate for dynamic querying, synthesis, and reuse.

The cumulative direction of the field is clear from the surveyed work. RKGs are being developed as ontology-anchored, identifier-rich, queryable infrastructures that combine formal graph models, shared vocabularies, automated IE, community curation, FAIR-aware tooling, and increasingly neuro-symbolic interfaces. A plausible implication is that future progress will depend less on any single extraction or reasoning method than on sustained integration of representation, provenance, validation, and curator-centered governance across the full lifecycle of scholarly knowledge.