Ontology-Driven Conceptual Navigation

Updated 14 July 2026

Ontology-driven conceptual navigation is an approach that uses explicit ontological structures, like taxonomies and semantic graphs, to guide knowledge exploration.
It leverages techniques such as dynamic path construction, conceptual grounding, and lateral navigation to enhance semantic search, visualization, and adaptive learning.
Applied across science, education, and healthcare, this paradigm improves resource retrieval and deepens understanding of complex semantic relationships.

Searching arXiv for recent and foundational papers on ontology-driven conceptual navigation and ontology-based knowledge exploration. Ontology-driven conceptual navigation is a mode of knowledge exploration in which navigation is organized by explicit conceptual structures—ontologies, taxonomies, concept hierarchies, conceptual graphs, or other formally specified semantic systems—rather than by raw keywords, fixed hyperlinks, or undifferentiated result lists. In this paradigm, concepts are treated as primary navigational objects; documents, entities, and instances are connected to those concepts; and movement through an information space proceeds vertically through broader–narrower structure, laterally through semantic relatedness, or outward from concepts to relevant resources. Across the literature, this idea appears in Web navigation, semantic search, digital libraries, scientific knowledge systems, ontology visualization, educational systems, and knowledge graph interfaces, with distinct emphases on ontological grounding, runtime reasoning, conceptual indexing, dynamic path construction, and large-scale graph-based exploration (0705.1886, Shen et al., 2018, Zimmermann et al., 2024).

1. Foundations and definitions

Ontology-driven conceptual navigation is distinguished from both hyperlink navigation and conventional information retrieval. In hyperlink-based systems, links are pre-authored into documents; in keyword-based retrieval, users submit terms and receive ranked document lists. Conceptual navigation instead relies on formal conceptual descriptions of resources and on ontologies that specify concepts and relations, allowing links and paths to be computed from semantic structure rather than authored in advance (0705.1886). A related motivation appears in ontology-based search systems, where keyword matching and ranking algorithms such as PageRank, HITS, SALSA, and Hilltop are criticized for failing to account for semantic relations, motivating ontology-based retrieval and inference (Mukhopadhyay et al., 2013).

Within this literature, an ontology is consistently treated as an explicit conceptual structure describing a domain’s concepts and the relations that hold among them. In some systems it is a class-and-property structure in OWL or OWL-Lite, with classes, individuals, object properties, data properties, rdfs:subClassOf, rdfs:domain, rdfs:range, and owl:disjointWith used to define taxonomies and constraints (Mukhopadhyay et al., 2013). In others, it is a domain ontology expressed as conceptual graphs or in a conceptual language that constrains annotation and supports runtime navigation and pedagogical assembly (0705.1886). At Web scale, the ontology may take the form of a directed acyclic graph over fields of study, where concepts are grounded in Wikipedia entities and linked to publications by confidence-weighted concept–document edges (Shen et al., 2018).

A recurring distinction is between ontology-supported and ontology-driven navigation. In ontology-supported navigation, the ontology normalizes indexing, constrains descriptions, and provides the conceptual substrate, while the navigation strategy is encoded mainly in the engine. In ontology-driven navigation, the ontology also participates in runtime decision-making, driving selection, ordering, and assembly strategies through domain knowledge and, in some cases, pedagogical or argumentative ontologies (0705.1886). This distinction is important because it separates systems that merely use ontologies as vocabularies from systems in which ontological structure actively governs the navigational process.

A further theoretical strand connects conceptual navigation with semantics and explanation. Ontological unpacking treats conceptual models, knowledge graphs, and symbolic specifications as truth-bearers whose truthmakers are entities and relations in an ontology. This suggests that navigation over symbolic structures can be grounded in explicit ontological commitments rather than surface syntax. A plausible implication is that ontology-driven navigation is not only a retrieval strategy but also a method for revealing the semantic commitments of a representation (Guizzardi et al., 2023).

The basic navigational substrate varies across systems, but several recurrent structures appear.

One family uses hierarchical concept systems. In a large-scale scientific knowledge system, the ontology is a six-level directed acyclic graph with levels $L0$ – $L5$ , where edges encode broader–narrower subsumption relations. The top two levels are manually curated, while lower levels are inferred automatically, yielding more than 200 thousand concepts and more than 1 million parent–child edges (Shen et al., 2018). In Ontoverse, the Topic Hierarchy Graph functions as the “navigational backbone,” organizing topics and subtopics for cartographic exploration; it can be derived from pre-existing ontologies, proprietary schemes, or hierarchical topic modeling (Zimmermann et al., 2024). In OntologyNavigator, the ACM Computing Classification System is transformed from a taxonomy into a richer ontology with hierarchical and relatedness relations, supporting visual navigation and cross-language retrieval (Kembellec et al., 2011).

A second family uses formal contexts and concept lattices. CKML treats formal contexts and concept lattices as equivalent data structures, enabling the same information to be navigated either as an incidence structure of objects and attributes or as a lattice of formal concepts ordered by specialization (Kent, 2018). Conceptual scaling then converts ontologically structured metadata into faceted conceptual spaces, allowing navigation by moving through lattice structure or switching among scales and facets (Kent, 2018).

A third family emphasizes graph-based conceptual descriptions. In Karina and related systems, resources are described by conceptual state vectors composed of weighted conceptual graphs of the form $[\text{SOURCE}] - [\text{PREDICATE}] - [\text{DESTINATION}]$ . Navigation proceeds by conceptual proximity, prerequisites, and related conceptual descriptions rather than by static links (0705.1886). In health-care semantic search, OWL-Lite ontologies encode classes, subclass relations, properties, and individuals, enabling browsing through disease classes, symptoms, and remedies (Mukhopadhyay et al., 2013).

A fourth family couples conceptual structure with instance-level or document-level grounding. In the Microsoft Academic Graph system, concepts are tied to publications through approximately 1B concept–publication pairs with scores, allowing movement from a paper to its fields-of-study, their parents, children, and related concepts (Shen et al., 2018). Ontoverse separates the Core Entity Graph, Topic Hierarchy Graph, and Topic Occupancy Graph, explicitly linking publications to topic nodes and allowing both topic-driven and similarity-driven exploration (Zimmermann et al., 2024). OnSET similarly uses ontologies and knowledge graphs to let users build prototype subgraphs and explore matching instances without SPARQL (Kantz et al., 11 Apr 2025).

Ontology-driven conceptual navigation typically combines three operations: conceptual grounding, path construction, and concept-to-resource linkage.

Conceptual grounding maps raw text, entities, or user expressions into ontology concepts. In scientific knowledge exploration, concepts are grounded in Wikipedia entities, each backed by a specific article, with the first paragraph used as simple representing text and richer extended representing text used for broader domains (Shen et al., 2018). In Ontoverse, named entities extracted from text are normalized to UMLS concept identifiers and filtered to biomedical ontologies such as HPO, DRUGBANK, MeSH, HGNC, NCI, and RXNORM (Zimmermann et al., 2024). In OntologyNavigator, bilingual node labels, user-provided alternative translations, and implicit descriptors such as “C++” or “C#” map user vocabulary into ontology nodes (Kembellec et al., 2011). In the Twitter disease-incidence system, tweets are converted into a Concept Normal Form by replacing lexical items with ontology concept labels or POS tags, producing a fixed conceptual alphabet for learning and classification (Magumba et al., 2016).

Path construction may be static, inferred, or computed at runtime. In subsumption-based scientific ontologies, the hierarchy is inferred using confidence-weighted concept–document co-occurrence, with relative coverage used to decide when one concept becomes the child of another (Shen et al., 2018). In Karina, backward conceptual navigation begins with an objective conceptual state vector, selects the resource with maximal conceptual proximity, inserts prerequisites as new objectives, and organizes resources by prerequisite structure, with time constraints handled through backtracking (0705.1886). In Sybil, a pedagogical ontology supplies rules such as “Explanation precedes Example,” and a domain ontology decomposes concepts into subgoals, so the resulting path is both conceptually and pedagogically structured (0705.1886). In health-care search, DL reasoners classify the ontology, compute inferred hierarchies, and support navigation over asserted and inferred class structure (Mukhopadhyay et al., 2013).

Concept-to-resource linkage connects conceptual nodes to retrievable items. In the Microsoft Academic Graph system, each publication receives multiple fields-of-study labels via cosine similarity between concept and publication vectors, using both text and graph structure, including citations, references, and venues (Shen et al., 2018). In OnSET, the user’s prototype graph is translated into SPARQL and matched against subgraph instances in the knowledge graph, with the ontology constraining valid expansions (Kantz et al., 11 Apr 2025). In OntologyNavigator, the current ontology node determines meta-queries to internal and external bibliographic services, so concept navigation directly yields article lists and bibliographic material (Kembellec et al., 2011). In the health-care ontology system, URLs are represented as individuals associated with disease and remedy concepts, allowing concept-level retrieval (Mukhopadhyay et al., 2013).

Several systems also support lateral navigation. Related concepts may arise from shared parents, shared tagged papers, descriptor co-occurrence, or explicit “related” displays (Shen et al., 2018, Kembellec et al., 2011). Ontoverse uses multi-topic occupancy and clone nodes so a single publication can appear in multiple topic regions, supporting navigation across conceptual neighborhoods and hidden topic bridges (Zimmermann et al., 2024). This suggests that lateral movement is often a crucial complement to taxonomic browsing.

4. Major system paradigms

Different research traditions instantiate ontology-driven conceptual navigation in distinct ways.

Concept-hierarchy systems

The Web-scale scientific knowledge system is a canonical example of hierarchy-centered navigation. It identifies approximately 228K fields of study from over 5M English Wikipedia entities, tags concepts to hundreds of millions of publications, and builds a six-level DAG embedded in MAG (Shen et al., 2018). Users can move top-down from broad disciplines to fine-grained topics, bottom-up from papers to concepts, and laterally through related fields. The system supports conceptual navigation because the ontology, the concept–document graph, and the document graph are integrated in one structure (Shen et al., 2018).

Ontoverse adopts a related but more explicitly visual approach. Its Core Entity Graph captures publication similarity based on shared ontology-grounded concepts; the Topic Hierarchy Graph provides the navigational backbone; and the Topic Occupancy Graph handles multi-topic occupancy via entity clones. Circle packing and topographic metaphor expose hierarchical topic structure as a cartographic surface with semantic zooming (Zimmermann et al., 2024).

Karina and Sybil represent a different paradigm in which navigation paths are computed from conceptual descriptions, goals, and constraints. Resources are self-contained; semantic links are not stored inside them; instead, meta-descriptions and ontologies allow a conceptual engine to build paths dynamically (0705.1886). In Karina, the user’s initial knowledge, learning objective, and available time drive backward conceptual navigation. In Sybil, a pedagogic ontology and a domain ontology jointly determine how to explain a concept such as sonata form by top-down exposition, explanation, and example sequencing (0705.1886). Here ontology-driven navigation is tightly coupled to adaptive pedagogy.

Ontology-based search and retrieval

Health-care semantic search with Protégé-OWL and OWL-Lite treats ontologies as the basis for semantic indexing and inference over web pages. Query preprocessing maps lexical items to ontology classes or properties; a DL reasoner computes the inferred class hierarchy; and results can include conceptually related items such as migraine pages in response to “headache medicine” (Mukhopadhyay et al., 2013). Ontology-driven personalized information retrieval for XML documents extends this pattern by representing documents, queries, and user profiles as vectors of ontology concepts, weighting deeper ontology concepts more strongly, and computing relevance through semantic similarity in concept space (Iddir et al., 22 Mar 2026).

Knowledge graph and visual exploration systems

OnSET targets users without prior knowledge of SPARQL or the ontology. Topic modeling over class descriptions produces human-readable entry topics, semantic search suggests start links and graph expansions, and prototype graphs are turned into executable subgraph queries whose instance matches are shown as small multiples (Kantz et al., 11 Apr 2025). OntView focuses on ontology exploration itself: it reasones over OWL 2 ontologies, visualizes inferred TBox structure including GCIs and anonymous classes, and supports summarization, fragment restriction, and semantics-preserving hide/show operations (Bobed et al., 18 Jul 2025). FuzzyVis advances this line by allowing users to visually compose new concepts with conjunction, disjunction, and negation, then retrieve nearby ontology concepts via fuzzy ontology embeddings and cosine similarity (Zhurov et al., 11 Aug 2025).

Some systems emphasize the interpretability of the ontology rather than only retrieval performance. Contextual verbalization pairs diagram elements with controlled natural language explanations of the underlying OWL axioms, making ontology diagrams self-explanatory at the point of interaction (Liepiņš et al., 2016). OntView shows the actual inferred knowledge rather than only asserted structure, following a “What you see is what you meant” paradigm and explicitly visualizing GCIs (Bobed et al., 18 Jul 2025). These systems suggest that ontology-driven navigation often requires explanation layers so users can interpret what a concept, relation, or restriction actually means.

5. Formal models and computational techniques

Several formal mechanisms recur across the literature.

Subsumption is one of the most common. In the scientific knowledge system, hierarchy construction extends the classical idea that a term $x$ subsumes $y$ when almost all documents containing $y$ also contain $x$ , using weighted co-occurrence over concept–document scores rather than raw frequencies (Shen et al., 2018). This yields a DAG rather than a strict tree and supports multiple parentage.

Conceptual state vectors and conceptual proximity support computed navigation in Karina. A conceptual state vector is a symbolic sum of weighted conceptual graphs,

$\text{CSV} = \sum_i w_i \cdot CG_i, \quad w_i \in [0,1],$

and conceptual proximity measures how much of the current objective is covered by a resource’s content description (0705.1886). This underlies resource selection, prerequisite insertion, and course assembly.

Formal contexts and concept lattices provide another rigorous basis. A formal context is a triple $(G, M, I)$ , where $G$ is a set of objects, $L5$ 0 a set of attributes, and $L5$ 1 the incidence relation, and the resulting concept lattice orders formal concepts by specialization (Kent, 2018). This offers a mathematically precise model for navigation by refinement and generalization.

Vector-space semantics also appear repeatedly. The Microsoft Academic Graph system computes concept–paper confidence as cosine similarity between concept vectors and publication extended representing text vectors (Shen et al., 2018). Personalized XML retrieval represents text nodes, element nodes, queries, and user profiles as vectors of ontology concepts and uses cosine similarity for ranking (Iddir et al., 22 Mar 2026). FuzzyVis represents primitive concepts as fuzzy membership vectors and composes new query concepts by element-wise fuzzy operators, then uses cosine similarity to retrieve nearest ontology concepts (Zhurov et al., 11 Aug 2025).

Knowledge graph similarity is central to Ontoverse, where publication–publication similarity is the count of shared ontology-grounded concepts, with only edges above a threshold retained to keep the Core Entity Graph interpretable (Zimmermann et al., 2024). OnSET uses semantic embeddings for classes and links, ranking candidate expansions by similarity between user-entered phrases and ontology property descriptions (Kantz et al., 11 Apr 2025).

Reasoning is another major component. OWL-Lite health-care navigation uses FaCT++ or Pellet for classification and consistency checking (Mukhopadhyay et al., 2013). OntView uses Openllet and HermiT to compute inferred direct superclasses, equivalences, and anonymous expression placements (Bobed et al., 18 Jul 2025). Contextual verbalization relies on mapping each diagram element to its associated OWL axioms and rendering them in Attempto Controlled English (Liepiņš et al., 2016).

6. Applications and domains

Ontology-driven conceptual navigation has been applied across substantially different domains.

Scientific literature exploration is a central application. The Microsoft Academic Graph system enables conceptual movement through hundreds of millions of publications via a cross-domain scientific ontology (Shen et al., 2018). OntologyNavigator uses an ontology derived from the ACM Computing Classification System to help advanced learners and early researchers construct IT bibliographies, with bilingual navigation and links to ACM, DBLP, CiteSeer, and other sources (Kembellec et al., 2011). Ontoverse targets biomedical literature, grounding publications in domain ontologies and exposing the resulting structure through a map-like interface (Zimmermann et al., 2024).

Education and adaptive learning form another major domain. Karina dynamically builds courses from resource descriptions, learner objectives, and time constraints (0705.1886). Sybil uses classical music ontology plus pedagogical ontology to generate concept-driven teaching paths (0705.1886). These systems show that navigation can be toward learning outcomes rather than document retrieval alone.

Biomedical and health-care applications include disease and remedy search over crawled web pages (Mukhopadhyay et al., 2013), biomedical ontology exploration via fuzzy concept composition (Zhurov et al., 11 Aug 2025), and social-media disease incidence detection, where tweets are interpreted through an ontology of disease-language concepts such as symptoms, time, locality, magnitude, and negation (Magumba et al., 2016). In the Twitter setting, the ontology supports concept-level filtering across lexically divergent diseases by training models on concepts rather than words (Magumba et al., 2016).

Semi-structured document retrieval is represented by ontology-driven personalized XML search, where ontology structure, profile vectors, and XML tree distance jointly shape retrieval (Iddir et al., 22 Mar 2026). Knowledge representation and conceptual modeling are addressed through ontological unpacking and ontological components, which emphasize navigation across meaning structures and dependencies rather than only data objects (Guizzardi et al., 2023, Fabry et al., 27 Mar 2025).

A more unusual application appears in robotics and embodied AI. Analytic Ontology Templates encode geometric and kinematic concepts, and AOTNet maps partial point clouds of articulated objects into explicit analytic descriptions of parts, joints, and affordances, allowing an agent to operate at the conceptual level on novel categories (Sun et al., 2024). This suggests that ontology-driven conceptual navigation need not be limited to document or graph spaces; it can also apply to action-oriented concept spaces.

7. Limitations, controversies, and open directions

A persistent limitation concerns ontology quality and semantics. Subsumption-based hierarchy induction can yield type-inconsistent edges, as when “polycystic kidney disease” becomes a child of “kidney” because of asymmetric mention patterns in documents (Shen et al., 2018). The paper explicitly notes intransitivity and lack of type constraints, indicating that pure statistical subsumption is not identical to ontological correctness (Shen et al., 2018). A similar issue arises in data-driven topic hierarchies: when no curated ontology exists, systems such as Ontoverse rely on topic modeling, which may produce less interpretable categories (Zimmermann et al., 2024).

Another recurrent challenge is scalability versus comprehensibility. Large ontologies overwhelm users if shown naively. OntView addresses this through KCE, PageRank/RDFRank summaries, fragment restriction between two classes, and percentage-based expansion and collapse (Bobed et al., 18 Jul 2025). Contextual verbalization points out that diagrams alone are not self-explanatory and that users must otherwise learn the visual notation before understanding the ontology (Liepiņš et al., 2016). OnSET similarly assumes that users do not know the ontology or SPARQL and therefore requires topic modeling and semantic search as scaffolding (Kantz et al., 11 Apr 2025). This suggests that ontology-driven navigation systems require layered interfaces: overview, explanation, and interaction aids.

Semantic expressiveness is also a tension point. OWL-Lite health-care search deliberately restricts expressivity for tractable reasoning (Mukhopadhyay et al., 2013). FuzzyVis currently focuses mainly on concept-level constructors—intersection, union, and negation over concept names—and proposes richer role restrictions and relationship support as future work (Zhurov et al., 11 Aug 2025). Personalized XML retrieval uses a weighted concept-vector model rather than more expressive logical inference (Iddir et al., 22 Mar 2026). These choices indicate that many practical systems trade expressive completeness for interactional or computational tractability.

Evaluation remains uneven. Some systems report quantitative accuracy or retrieval metrics: concept discovery 94.75%, concept tagging 81.20%, and hierarchy building 78.00% in the scientific knowledge system (Shen et al., 2018); average F1 around 0.781 in personalized XML retrieval (Iddir et al., 22 Mar 2026); strong cross-disease robustness in ontology-driven Twitter classification (Magumba et al., 2016). Others emphasize prototypes, qualitative case studies, or architectural principles without controlled user studies, including Karina, Sybil, OntView, and contextual verbalization (0705.1886, Liepiņš et al., 2016, Bobed et al., 18 Jul 2025). A plausible implication is that the field still lacks a unified evaluation protocol for conceptual navigation quality, especially one that jointly measures correctness, discoverability, interpretability, and user learning.

A deeper controversy concerns what counts as meaning in ontology-driven navigation. Ontological unpacking argues that explanations should reveal the ontological commitments and truthmakers of symbolic models (Guizzardi et al., 2023). Ontological components propose that meaning be operationalized through entailment-based meaning specifications combining analytic statements and natural-language characterizations (Fabry et al., 27 Mar 2025). These approaches imply that navigation among concepts should perhaps be based not merely on taxonomic or similarity structure but on explicit meaning specifications. That direction remains theoretically rich but operationally demanding.

Future directions recur across the papers. They include richer relation types beyond subsumption, such as part-of, causal, or method-for (Shen et al., 2018); incorporation of KB type information to enforce hierarchy consistency (Shen et al., 2018); more advanced graph analytics and path-based exploration in map-like interfaces (Zimmermann et al., 2024); guided tours and inferred-axiom explanations for ontology visualizations (Liepiņš et al., 2016); personalization of concept ranking and navigation paths (Shen et al., 2018); stronger support for reuse and version-aware meaning navigation via ontological components (Fabry et al., 27 Mar 2025); and broader ABox-level navigation over reasoned ontologies (Bobed et al., 18 Jul 2025).

Taken together, these works show that ontology-driven conceptual navigation is not a single technique but a design space. Its unifying principle is that information spaces become navigable when concepts, relations, and ontological commitments are made explicit, operational, and connected to the resources or instances that users ultimately seek (0705.1886, Shen et al., 2018, Kantz et al., 11 Apr 2025).