Automated Ontology Generation (AOG)
- Automated Ontology Generation (AOG) is a process that automatically converts structured and unstructured data into formal semantic models like OWL ontologies and knowledge graphs.
- It employs multiple pipelines—from XML-Schema mapping to natural language processing and supervised classification—to extract and structure concepts, relationships, and axioms.
- Human oversight and adherence to interoperability standards ensure that the generated ontologies are refined, verifiable, and ready for downstream applications such as retrieval and question answering.
Automated Ontology Generation (AOG) denotes the automatic or semi-automatic production of ontology artifacts from structured and unstructured sources. Across the literature, the target artifacts include OWL classes and properties, Description Logic axioms, fuzzy ontologies, ontology graph scaffolds, and knowledge graphs populated with instances. The source material ranges from XML data sources and registry metadata to product reviews, competency questions, user stories, software engineering standards, scholarly publications, and technical documents (Yahia et al., 2012, Thomas, 2015, Lippolis et al., 7 Mar 2025).
1. Problem scope, source modalities, and target artifacts
AOG is not a single algorithmic task but a family of transformations from source representations to formal semantic structures. In structured settings, the source can be XML documents and XML Schemas; in corpus-driven settings, it can be short unstructured text, domain-specific verbatims, or curated subject fields; in recent LLM-supported settings, the source can be competency questions, user stories, GitHub issues, or long technical documents. The outputs likewise vary: OWL-DL ontologies, ontologies in the Description Logic fragment SHOQ(D), ontology graph scaffolds of the form , and fuzzy ontologies of the form (Yahia et al., 2012, Dasgupta et al., 2018, Yue, 29 Aug 2025, Touzi et al., 2013).
| Source modality | Representative procedure | Output form |
|---|---|---|
| XML and XSD | XML-Schema Graph generation and mapping rules | OWL ontology |
| Registry subject fields | tokenization, filtering, and ontology merging | OWL classes merged with IVOAT |
| Competency questions and user stories | few-shot or iterative prompting | Turtle/OWL ontology fragments |
| Long technical documents | triplet extraction and graph construction | ontology scaffold or KG |
A recurrent distinction is between TBox generation and ABox population. One line of work generates the conceptual schema from competency questions and then constructs KGs using that ontology; another directly maps extracted entities, relationships, and attributes into populated graphs or RDF triples (Kommineni et al., 2024, Abolhasani et al., 2024). Another distinction concerns whether the objective is de novo ontology construction, ontology completion, or ontology refinement against a reference ontology. DRAGON-AI, for example, is explicitly a method for semi-automated ontology generation and completion, whereas some KG-centered frameworks emphasize correction, completion, and mapping to target ontologies (Toro et al., 2023, Elnagar et al., 2022).
The acronym is also overloaded. In computer vision, AOG can denote an And-Or Graph, not automated ontology generation. That usage refers to a four-layer semantic hierarchy of semantic parts, part templates, latent patterns, and neural units mined from convolutional networks, and is conceptually distinct from ontology engineering workflows (Zhang et al., 2018).
2. Rule-based, schema-driven, and feature-engineered methods
A major early strand of AOG works by lifting structured or weakly structured sources into formal ontology languages through deterministic mappings. In XML-based generation, the pipeline is: XML to XML-Schema conversion with Trang; XML Schema parsing and object model extraction with XSOM; graph representation as an XML-Schema Graph (XSG) using JUNG; and OWL ontology generation using Jena. The mapping rules are explicit: global named complex types become OWL classes, XML Schema extension and restriction become rdfs:subClassOf, element–subelement relations become OWL object properties, and simple type elements and attributes become datatype properties (Yahia et al., 2012).
A second strand uses linguistic normalization and hand-specified semantic translation rules. DLOL first normalizes English factual non-negative IS-A sentences with a Normalized Sentence Structure and then translates them into SHOQ(D) through a Description Logics-based Ontology Learner pipeline. This is intended to generate both TBox and ABox knowledge, including modifiers, quantifiers, and non-trivial IS-A patterns. The same tradition appears in the radiological protection literature, where semi-automatic ontology construction proceeds through text annotation with part-of-speech tags, revelation of significant linguistic structures and forming the templates, search of text fragments corresponding to these templates, and a basic ontology instantiation process (Dasgupta et al., 2018, Orobinska, 2014).
A third strand uses feature engineering and supervised classification for concept identification and typing. In automotive complaint and repair verbatim data, candidate collocates are extracted after misspelling correction, run-on word correction, extra whitespace removal, and abbreviation disambiguation. A two-stage classification system then distinguishes concepts from irrelevant collocates and assigns discovered concepts to ontology concept types. The reported classifiers are Random Forest models using POS, word2vec, context, polysemy, and ontology-based features (Xu et al., 2019).
Formal Concept Analysis and clustering provide yet another route. In the Fuzzy Ontology of Data Mining approach, fuzzy clustering is performed first, then an -cut converts fuzzy memberships into a binary context suitable for FCA, and ontology concepts and relations are generated from a fuzzy lattice. The stated novelty is the additional degree of granularity introduced by defining the ontology between classes resulting from a preliminary classification on the data rather than at the raw instance level (Touzi et al., 2013).
3. Corpus mining, ontology merging, and graph transformation
Corpus-based AOG often begins with aggressive regularization of input terms before any ontology construction occurs. A clear example is the Virtual Observatory registry workflow. Its primary input is the subject field from approximately 30,000 resources, and about 95% of these subject entries are single nouns or noun phrases, with about 5% full sentences. Sentence-based fields are filtered out, leaving a corpus of about 1,100 unique text instances; tokenization, filtering through a local astronomical dictionary, Wikipedia, and WordNet reduce this to about 450 validated subject concepts. The generated classes are initially flat under owl:Thing, then merged against the IVOA Thesaurus transformed from SKOS into OWL. Direct lexical matching and indirect hypernym matching are both used, and the reported structural score is 0.32 over about 450 subject concepts (Thomas, 2015).
Product review mining shows a different corpus-driven pattern: ontology extraction in the form of meronomies. The pipeline begins with a small handcrafted ontology for only 5 products with up to 54 (sub-)features, then uses distant supervision to label review sentences. It extracts frequent noun phrases, classifies aspects with a masked BERT classifier, groups terms into synsets via word2vec and Equidistant Nodes Clustering, classifies relations with a second masked BERT classifier, and constructs a hierarchical meronomy tree from a relation matrix. The reported averages are 76.9% precision, 54.9% relative recall, and 63.2% macro F1, with 70% aspect coverage and 15.8% relation coverage on an Amazon Q&A dataset (Oksanen et al., 2021).
Graph-centric frameworks move from text to preliminary knowledge graphs before ontology projection. One fully automatic, domain-independent framework first generates a KG from unstructured text, then refines it through anomaly exclusion, correctness checking, and completion. Triples with confidence below 0.3 are removed; borderline cases with confidence 0.3–0.5 are processed with Local Outlier Factor; logical correction uses disjointness axioms from DOLCE and association rule mining; and missing links are predicted with ComplEx embeddings. A final mapping phase removes triples unsupported by the target ontology’s domain and range constraints (Elnagar et al., 2022).
A related but practically oriented transformation starts not from free text alone but from a property graph. In the health domain, FAERS data are loaded into Neo4j, then a Python application uses the Neo4j driver and rdflib to extract schema and instance data, create classes and properties, assert domain and range constraints, add subclass relations and OWL restrictions, and serialize the ontology to .owl. The stated purpose is to lower the barrier posed by direct description logic authoring while preserving ontology classes and their axioms (Gadusu et al., 25 Jun 2025).
4. LLM-supported AOG and ontology drafting from requirements
LLM-supported AOG introduces a different decomposition of the problem. Instead of relying only on pattern libraries or classifiers, these methods use controlled prompting, retrieval, and iterative validation. DRAGON-AI is a representative completion-oriented system: ontology terms, GitHub issues, and other sources are embedded and stored in ChromaDB; maximal marginal relevance is used during retrieval; a partial ontology object is completed by an LLM as a JSON object; and post-processing removes relationships to non-existent terms. The system generates both textual definitions and logical components such as relationships and logical definitions, and it can incorporate natural language instructions from GitHub issues (Toro et al., 2023).
A more restrictive line of work tries to align generation with upper ontologies or established ontological suites. “My Ontologist” uses a custom GPT with up to 20 PDF documents containing BFO specifications, BFO 2020 Turtle code, Common Core Ontologies, domain ontology guides, and other materials. Its third iteration adds 36 explicit rules, including adherence to the Aristotelian genus-differentia pattern, use of official class and property IRIs, and a ban on introducing unnecessary new classes and properties. The reported experience is mixed: modularization and structured rules improved BFO/CCO adherence, but GPT-4o altered the model’s behavior and increased property invention (Benson et al., 2024).
Requirement-driven generation from competency questions and user stories has become a major benchmark setting. One six-stage pipeline over scholarly publications uses competency questions to derive an ontology, applies Mixtral 8x7B for TBox creation, then answers competency questions by Retrieval-Augmented Generation and maps the answers into an RDF KG. The resulting ontology, “DLProv,” contains 45 classes, 41 relations, and 365 axioms. Human-in-the-loop is limited to CQ refinement and ground-truth annotation for evaluation (Kommineni et al., 2024).
More recent CQ-centered work compares prompting regimes directly. “Memoryless CQbyCQ” generates a separate ontology module for each CQ without seeing prior outputs, while “Ontogenia” incrementally merges previous ontology content and incorporates Ontology Design Patterns, labels, comments, inverse relationships, individuals, explanations, and test cases. On a benchmark of 10 ontologies with 100 distinct CQs and 29 different user stories, OpenAI o1-preview with Ontogenia is reported to produce ontologies of sufficient quality to meet the requirements of ontology engineers and to significantly outperform novice ontology engineers in modelling ability (Lippolis et al., 7 Mar 2025).
Cross-domain generalizability has been probed by generating ontology fragments from 95 curated competency questions and related user stories across six domains: Circular Economy, Music, Events, Microbe Habitat, Carbon and Nitrogen Cycling, and Water and Health. The models evaluated are DeepSeek and o1-preview, both described as equipped with reasoning capabilities, and the reported finding is that performance is remarkably consistent across all domains, indicating generalization of ontology generation tasks irrespective of domain (Lippolis et al., 24 Apr 2025).
Document-to-ontology pipelines also use LLMs for relation extraction rather than direct OWL authoring. In software engineering standards, the AOG workflow consists of document segmentation, candidate term mining with spaCy, LLM-based relation inference with strict JSON output at temperature , retry logic up to , term normalization, cross-section alignment, and ontology scaffold construction as . The study reports that a 7B open-source LLM is already competitive with OpenIE, with higher precision on both nodes and triples and comparable recall on stricter gold sets when (Yue, 29 Aug 2025).
Interactive pipelines make the ontology itself negotiable. OntoKGen explicitly presents ontology extraction as an adaptive, iterative Chain of Thought process in which the LLM suggests concepts, relationships, and properties, and the user confirms or modifies them before KG generation proceeds. Its ontology extraction instructions include concept identification, concept confirmation, relationship identification, relationship confirmation, and concept property identification, followed by KG generation and Cypher query construction for Neo4j (Abolhasani et al., 2024). A structurally similar but broader methodology leverages the agility of SAMOD and defines seven steps—Scenarios & Glossary Creation, Informal Competency Questions, Modelet Development, Test Case Generation, Model Refinement, Documentation Generation, and Feedback Integration—with LLMs supporting each step (Luyen et al., 15 Jan 2026).
5. Evaluation regimes and empirical results
AOG has no single dominant evaluation protocol. Different papers assess different artifacts: lexical coverage, structural alignment, logical utility, modeling adequacy, relation extraction quality, ontology pitfalls, or downstream QA performance. The consequence is a strongly multi-dimensional evaluation culture (Lippolis et al., 7 Mar 2025).
| Evaluation setting | Metric or protocol | Representative result |
|---|---|---|
| VO registry ontology | class match measure | 0.32 over about 450 subject concepts |
| DLOL | lexical accuracy and instance-based inference model | about 21% and 46% better than the best of Text2Onto, FRED, and LExO |
| Product meronomies | precision, relative recall, macro F1 | 76.9%, 54.9%, 63.2% |
| CQ-driven OWL drafting | OOPS!, modelled CQs, superfluous elements, expert review | Ontogenia + o1-preview reaches up to 100% adequately modelled CQs for certain stories |
| Domain-specific CQ study | Yes / No / No Minor manual categories | o1-preview: 83 Yes, 1 No, 8 No Minor; DeepSeek: 78 Yes, 5 No, 5 No Minor |
| SES triplet extraction | precision, recall, F1 over threshold sweep | higher precision than OpenIE; at , the matches are highly trustworthy |
The Virtual Observatory work uses a structural “Class Match Measure” adapted from Alani and Brewster, with score 1 for a direct match and 0.5 for an indirect match via a parent class. The score is
0
where 1 is the number of directly matched subject concepts, 2 the number matched via parent or hypernym, and 3 the total number of subject concepts used (Thomas, 2015).
DLOL, by contrast, combines lexical precision and recall with a novel instance-based inference model intended to measure structural and semantic preservation through inferred instances. Its reported comparative gains are about 21% on lexical recall and 46% on IIM-R over the best of Text2Onto, FRED, and LExO, and it is presented as uniquely able to formalize a wide range of trivial and non-trivial IS-A cases (Dasgupta et al., 2018).
LLM-era evaluations frequently mix automatic and human judgement. DRAGON-AI evaluates relationship generation with precision, recall, and F1 against recently added terms from ten major ontologies; for is-a relationships, GPT-4 is reported at precision 0.889, recall 0.44, and F1 0.588, while definition quality is scored by ontology editors and curators on 1–5 scales for accuracy, internal consistency, and overall utility. The paper also reports that GitHub issue integration improved definition accuracy for all tested LLMs (Toro et al., 2023).
CQ-driven ontology generation studies add ontology-engineering-specific criteria. One benchmark evaluates critical OOPS! pitfalls, the proportion of modelled CQs, superfluous classes and properties, and expert qualitative usability. Another study uses the categories Yes, No Minor, and No to assess whether a CQ is modelled at all. The SES relation-extraction work evaluates both nodes and triples with embedding-based one-to-one alignment and threshold sweeps, emphasizing that triple matching is stricter because errors in any of head, relation, or tail cause a miss (Lippolis et al., 7 Mar 2025, Lippolis et al., 24 Apr 2025, Yue, 29 Aug 2025).
6. Human oversight, interoperability, and downstream systems
A persistent misconception is that AOG now yields universally correct ontologies without expert intervention. The literature does not support that conclusion. OntoKGen states that there is no universally correct ontology because it is inherently based on the user’s preferences; users confirm, refine, or augment suggested concepts, relationships, and properties. DRAGON-AI concludes that expert curators and ontology editors should drive the ontology generation process. The scholarly-publication pipeline recommends a human-in-the-loop approach to evaluate automatically generated KGs. “My Ontologist” likewise shows that even highly constrained BFO-based generation remains sensitive to ambiguity, model updates, and hallucinated properties (Abolhasani et al., 2024, Toro et al., 2023, Kommineni et al., 2024, Benson et al., 2024).
Interoperability is a central design objective. Several pipelines target standards-based outputs such as OWL, RDF, Turtle, Cypher, or Neo4j property graphs. The FAERS workflow uses Python and rdflib to bridge Neo4j and OWL, with automatic generation of classes, properties, domain and range constraints, subclass axioms, and OWL restrictions. Other pipelines explicitly reuse foundational ontologies or standards, including PROV-O, BFO, Common Core Ontologies, IVOAT, and SKOS-to-OWL transformations (Gadusu et al., 25 Jun 2025, Kommineni et al., 2024, Benson et al., 2024, Thomas, 2015).
AOG is also increasingly embedded in downstream retrieval and question-answering systems. OG-RAG uses domain-specific ontologies to ground retrieval in a hypergraph of factual blocks and reports a 55% increase in recall of accurate facts, a 40% improvement in response correctness, 30% faster attribution of responses to context, and a 27% boost in fact-based reasoning accuracy across four different LLMs. OntoRAG, which derives ontologies from unstructured electrical relay documents through web scraping, PDF parsing, hybrid chunking, information extraction, knowledge graph construction, and ontology creation, reports a comprehensiveness win rate of 85% against vector RAG and 75% against GraphRAG’s best configuration (Sharma et al., 2024, Tiwari et al., 31 May 2025).
This suggests that AOG is increasingly treated not only as ontology authoring but also as infrastructure for KG construction, ontology-grounded retrieval, and domain-specific reasoning. At the same time, the corpus shows that reliability still depends on post-processing, ontology alignment, expert review, and explicit validation criteria rather than on raw generation alone (Lippolis et al., 7 Mar 2025, Luyen et al., 15 Jan 2026).
