Semantic-KG: Semantic Knowledge Graphs

• Semantic-KG is a semantically enriched knowledge graph where nodes and edges carry explicit meanings defined by ontologies such as RDFS or OWL.
• The construction paradigm leverages controlled triple extraction with iterative components like Generator, Verifier, and Pruner to ensure precision and semantic coherence.
• Applications span from optimized query processing and semantic similarity benchmarks to enhanced reasoning in recommendation systems and downstream analytics.

In the literature considered here, Semantic-KG denotes a knowledge graph whose nodes and edges are endowed with explicit semantics through schema, ontology, or other semantic control layers, and whose use extends beyond storage to construction, querying, reasoning, evaluation, and downstream systems. A canonical formalization treats a Semantic-KG as a directed, labeled graph with RDF triple-edges $E\subseteq V\times P\times (V\cup L)$ together with a schema $\Omega$ comprising class hierarchies, property hierarchies, domain and range constraints, and possibly OWL axioms; related work also characterizes semantically enriched representations as those integrating both ABox assertions and TBox information into a single continuous or symbolic framework (Khan, 2023, Alam et al., 2023).

1. Definition and conceptual scope

A Semantic-KG is defined as a directed, labeled graph in which nodes represent entities or concepts and edges represent semantically typed relations enriched by a formal ontology such as RDFS or OWL. In RDF terms, semantics arise from atomic triples $\langle s,p,o\rangle$, ontological axioms such as rdfs:subClassOf and rdfs:domain/range, and the induced entailments over class-instance and relation structure (Khan, 2023). In a closely related formulation, “Knowledge Graph (KG)” and “Semantic Knowledge Graph (Semantic-KG)” are used interchangeably for graphs whose nodes denote real-world entities or concepts, whose typed edges are drawn from an ontology, and whose nodes may carry literal attributes and class memberships (Bhatt et al., 2020).

This scope is broader than graph storage alone. In construction settings, semantics govern which triples are extracted, how they are verified, and which entities remain expandable in subsequent iterations. In representation learning, semantics appear as type constraints, domain/range restrictions, subclass relations, relation text, or description-logic axioms. In interaction settings, semantics determine query interpretation, task grounding, closure assumptions, and whether graph answers are actionable for a specific agent (Chen et al., 2024, Arun et al., 26 May 2025, Payne et al., 18 May 2026).

A further granularity is the named subgraph. Semantic Units define a semantically significant subgraph within a larger RDF graph as a pair $(u,S)$, where $u$ is a URI identifying the unit and $S=(V_S,E_S)$ is a subgraph whose triples jointly express one unit of meaning, such as a measurement, a link, or an authorship statement. This moves Semantic-KG from whole-graph semantics to explicitly packaged local semantics (Mustafa, 27 Nov 2025).

2. Construction paradigms

A major line of work treats Semantic-KG construction as controlled triple extraction from corpora. SAC-KG formulates domain KG construction as extracting factual triples $(h,r,t)\in T\subseteq V\times R\times V$ and expanding them breadth-first over levels $k=0,\dots,K$ from a seed entity set $E^{(0)}$, with the final graph given by $G=\bigcup_{k=0}^{K}\widetilde S^{(k)}$. Its iterative loop consists of a Generator, which retrieves domain context and exemplar triples and prompts an LLM to propose candidate triples; a Verifier, which applies rule-based checks and LLM reprompts to correct or discard erroneous triples; and a Pruner, a fine-tuned T5 binary classifier that marks tails as “growing” or “pruned” for the next level (Chen et al., 2024).

Other construction pipelines introduce intermediate semantic structures before triple formation. SocraticKG uses question–answer pairs as an explicit semantic scaffold, defining the overall mapping as

$\Omega$0

where $\Omega$1 generates 5W1H-guided QA pairs from a document, $\Omega$2 extracts atomic facts from those pairs, and $\Omega$3 canonicalizes them into triples. Its motivation is to address a coverage–connectivity trade-off in direct LLM-based extraction by unfolding document semantics before graph consolidation (Choi et al., 15 Jan 2026).

Document-centric Semantic-KG construction also appears in ontology-heavy domains. For industrial standards, one framework models document layout as a rooted directed acyclic graph $\Omega$4, recursively decomposes sentences and tables into atomic propositions using propositional logic, defines an OWL-style ontology with classes such as AtomicProposition, NumericConstraint, Scope, and Exception, and then performs LLM-based triple extraction followed by normalization and pruning before merging the hierarchy and extracted triples into Neo4j (Park et al., 9 Dec 2025). A different workflow for educational content transforms an expert-curated five-level hierarchy of learning objects into a KG by text mining titles and descriptions, retaining the original is_subClassOf or part_of relations and adding has_semantic_relation_to edges when title or description similarity exceeds tuned thresholds (Abu-Rasheed et al., 2024).

Low-resource and multilingual settings require additional semantic filtering. BanglaAutoKG begins with multilingual LLM extraction of Bengali entities and relations, uses a Bangla-to-English translation dictionary plus pre-trained BERT embeddings to initialize node features, applies a self-supervised denoising GNN, and then combines topological and Chebyshev-polynomial spectral filters with an attention-based semantic filter to prune low-quality edges (Wasi et al., 2024).

3. Precision control, semantic filtering, and graph quality

Semantic-KG construction is not reducible to extraction volume; the cited systems make precision-control explicit. In SAC-KG, the Verifier detects three error classes—Quantity Check, Format Check, and Conflict Check—and reprompts the LLM with error-specific instructions. The Pruner then limits uncontrolled expansion by allowing only semantically meaningful tails to become seeds for the next level. In the reported rice-domain experiments, SAC-KG constructs a KG at the scale of over one million nodes, reaches Precision = 89.32\%, Domain Specificity = 81.25\%, and exceeds PIVE’s 64.48\% precision by more than 20 percentage points; ablations over three levels show that removing the open KG retriever, domain-text retriever, Verifier, or Pruner reduces precision by up to 12 pp and causes error propagation, especially without the Pruner (Chen et al., 2024).

Graph quality is evaluated differently when the construction objective is contextual linkage rather than maximal extraction. In the personalized-learning KG, semantic completion changes the graph from Average Degree Centrality 1.079 to 2.262, increases the number of detected communities from 253 to 541, decreases Modularity $\Omega$5 from 0.779 to 0.636, reduces Weakly Connected Components from 63 to 35, and raises Average Betweenness Centrality from 1.57 to 15.1. In a random sample of 240 new semantic relations, 79\% were judged at least as meaningful as expert-defined ones under the paper’s similarity criterion (Abu-Rasheed et al., 2024).

For BanglaAutoKG, semantic coherence is evaluated by the Average Semantic Feature Alignment Score (A-SFAS). The ablation reported for two Bangla text genres shows A-SFAS ≈ 0.45–0.53 without Feature Denoising or Semantic Filtering, ≈ 0.53–0.62 with Feature Denoising only, and ≈ 0.89–0.91 when Semantic Filtering is enabled, with the combination of Feature Denoising and Semantic Filtering also yielding ≈ 0.89–0.91 (Wasi et al., 2024).

In industrial standards, quality control targets reasoning usefulness rather than only extraction fidelity. The ontology-aware KG-RAG framework raises average F1 from 0.277 for Text-only RAG and 0.304 for existing KG-RAG to 0.454, described as +64\%. The strongest gain is on table questions (+93.7\%), with multi-hop (+59.7\%) also reported, and toxic-clause detection F1 up to 0.91 with Recall 0.926 (Park et al., 9 Dec 2025).

4. Semantic-aware representations and evaluation

A central research question is whether a KG representation is genuinely semantic-aware or only structurally predictive. One formulation defines a semantically enriched KG embedding as one that integrates not only the ABox but also TBox information—schema, type hierarchies, and description-logic axioms—into a single vector space. This line of work argues that domain/range constraints, subsumption, and ontological axioms reduce implausible completions, improve generalization for sparse entities, and can enforce logical coherence (Alam et al., 2023).

Evaluation research shows that rank-based metrics alone are insufficient for this purpose. Sem@K measures the fraction of top-$\Omega$6 predicted entities that satisfy domain/range constraints, with variants for schema-defined graphs, schemaless graphs, and class hierarchies. The reported experiments show that conclusions can diverge sharply from MRR and Hits@K: on FB15K237-ET, for example, ComplEx has MRR≈0.267 and Sem@10[base]≈0.761, TransE has MRR≈0.273 and Sem@10[base]≈0.949, and CompGCN reaches MRR≈0.337 and Sem@10[base]≈0.992. The paper also notes that Sem@K is not monotonic in $\Omega$7 (Hubert et al., 2023).

Schema-first transfer methods make semantics an initialization prior rather than a post hoc constraint. MASCHInE builds protographs from domain/range axioms and subclass structure, trains an off-the-shelf KG embedding model on that schema-derived graph, and transfers the resulting embeddings to the original KG. Reported results indicate that rank-based metrics remain on par with vanilla training, while Sem@K climbs from ∼0.97→0.99 on YAGO14k and DBpedia77k under P2, and entity clustering and node classification improve substantially (Hubert et al., 2023).

A separate misconception is that graph structure alone is enough for fully inductive reasoning. SEMMA addresses this by combining a structural relation graph with an LLM-enriched textual relation graph and fusing the two with a small MLP. Over 54 diverse KGs, the paper reports ULTRA: MRR ≈ 0.365 ± 0.003, Hits@10 ≈ 0.519 ± 0.004 versus SEMMA: MRR ≈ 0.377, Hits@10 ≈ 0.529, with statistically significant gains. In the stricter unseen-relation-vocabulary regime, SEMMA is described as nearly 2× better than ULTRA, including FBNELL: 0.025→0.058 MRR and Metafam: 0.098→0.180 MRR (Arun et al., 26 May 2025).

5. Querying, affordances, and human interaction

Semantic-KG querying spans formal graph languages, neural retrieval, and agent-oriented metadata. In the classical RDF setting, a Basic Graph Pattern $\Omega$8 is evaluated over a graph $\Omega$9 as

$\langle s,p,o\rangle$0

with joins defined over compatible mappings. The same survey situates Gremlin traversals, vector-based query processing, and KG-QA pipelines comprising entity linking, relation detection, semantic parsing, query planning, and answer ranking (Khan, 2023).

Domain-specific querying illustrates how Semantic-KG functions as a retrieval substrate rather than merely a repository. A climate-science KG defines node types such as Model, Dataset, Location, Teleconnection, Weather_Event, Paper, Variable, and Scenario, with relationship types such as MENTIONS, EVALUATED_WITH, and INFLUENCES. It supports Cypher questions such as which models have been validated in a region, which datasets are most used for Arctic validation, or which papers mention ENSO in South America, and the paper explicitly frames the KG as a component in LLM-based RAG for transparent scientific discovery (Adamu et al., 12 Sep 2025).

For agent planning, the notion of semantic capability is formalized by the Agentic Affordance Profile. The profile packages four dimensions:

$\langle s,p,o\rangle$1

where $\langle s,p,o\rangle$2 captures schema expressivity and compliance, $\langle s,p,o\rangle$3 agentic discoverability, $\langle s,p,o\rangle$4 task-relative grounding, and $\langle s,p,o\rangle$5 epistemic trust scope. This semantic layer is positioned above VoID and DCAT so that agents can select, compose, and diagnose KGs without exploratory probing at runtime (Payne et al., 18 May 2026).

Human-facing interaction raises a different issue: many technically valid triples are not semantically significant to end users. Semantic Units address this by packaging named, semantically significant subgraphs. In the Biodiversity Exploratories KG, the reported effect is that for known-resource queries, SU-based SPARQL becomes shorter and more focused, and the resulting visualizations are less cluttered; the same work also notes that for resource-discovery queries, SU-based queries still require mixing in the data graph and are as complex as conventional queries (Mustafa, 27 Nov 2025).

6. Applications, benchmark generation, and adjacent uses

Semantic-KG techniques appear in applications well beyond canonical KG completion. In end-to-end semantic communication, a KG-LLM semantic communication pipeline first extracts RDF triples from a sentence, serializes them, encodes them with T5-small, decodes them on the receiver side, and refines the text with BERT. The reported system reduces transmitted text volume by 30\% through KG-based compression and achieves 84\% semantic similarity between original and received messages, compared with 63\% for DeepSC; under 2–10 dB SNR, the detailed results show CR≈0.30 and semantic similarity rising from ≈0.58 to ≈0.84 (Salehi et al., 10 Apr 2025).

The term also names a benchmark-generation pipeline for semantic similarity evaluation. In that work, a KG is sampled into subgraphs, perturbed by node-removal, node-replacement, edge-removal, or edge-replacement, converted by an LLM into natural-language statements, and then used to construct positive and negative semantic-similarity pairs. The benchmark spans general knowledge, biomedicine, finance, biology and reports that no method for semantic similarity is consistently superior across perturbation types and domains. A manual spot-check found 99\% correctness and 75\% of statements judged “linguistically natural” (Wei et al., 25 Nov 2025).

Adjacent machine-learning literature uses Semantic-KG as an augmentation layer for input semantics. One surveyed framework enriches recommendation and community-detection systems by linking inputs to a KG, retrieving neighborhoods, extracting semantic features, and fusing them into downstream models; the reported gains include F1 ↑ 6.5\% in sentiment analysis, AUC: +2–3\% and CTR: +3–5\% for DKN-style news recommendation, Recall@20 ↑ 3–6\% and NDCG@20 ↑ 4–7\% for collaborative filtering, and modularity by 8–12\% for community detection (Bhatt et al., 2020). A related contextual LLM, CKG, augments tokens with one-hop DBpedia neighbors and reports SQuAD 89.2, compared with SAN 84.4, ELMo 85.8, and BERT$\langle s,p,o\rangle$6 88.5 (Tang et al., 2022).

Taken together, these results indicate that Semantic-KG is not a single architecture but a family of semantically constrained graph formalisms and workflows. The recurring technical themes are ontology-backed typing, explicit semantic control during construction, semantics-aware evaluation beyond rank metrics, and interaction layers that make graph meaning operational for reasoning systems, retrieval pipelines, and human users.