Ontology Expansion Techniques

• Ontology expansion techniques are systematic methods that add new concepts, relations, or axioms to existing ontologies, improving completeness and addressing domain drift.
• They combine deep learning approaches like NLI and GNN models with statistical pattern mining and modular extensions to ensure precision, scalability, and interoperability.
• Applications span diverse domains such as conversational AI, smart city infrastructures, and geospatial systems, facilitating robust and adaptable knowledge representations.

Ontology expansion techniques encompass the systematic addition of new concepts, relations, or axioms to existing ontologies, addressing incompleteness or domain drift in knowledge representations. The field includes data-driven statistical approaches, semantically motivated pattern mining, language-model-based inference, algebraic merging, and domain-driven modular extensions. The design and evaluation of ontology expansion methods integrate both empirical and formal considerations, including precision of enrichment, coverage, scalability, and compatibility with downstream applications.

1. Supervised and Deep Learning Approaches

Modern ontology expansion heavily employs supervised learning and deep architectures, especially transformers and LLMs. In the ontology completion paradigm, new subsumption axioms (e.g., $C \sqsubseteq D$) are predicted using two main classes of models:

• Natural Language Inference (NLI) Models: Each candidate inclusion is verbalized into natural language (e.g., via a rule-based verbalizer such that $$X \equiv \text{Biologist} \sqcap \exists \text{livesIn}.\text{UK}$$ becomes "a biologist who lives in the UK"). The input is encoded as $$[CLS]~\text{verbalize}(X)~[SEP]~\text{verbalize}(Y)~[SEP]$$ for a candidate rule $X \sqsubseteq Y$. A transformer encoder (e.g., RoBERTa, Llama2) encodes this, and an entailment head produces $s_{NLI}(r) = \sigma(w^\top h_r + b)$. Training is via binary cross-entropy loss over labeled inclusion/exclusion examples (Li et al., 2024).
• Concept Embedding with Graph Neural Networks (GNNs): Atomic concepts are embedded ($\mathbf{v}_C$), optionally via contextual mention averaging or bi-encoders. A concept co-occurrence graph is constructed, and embeddings are contextualized using GCN, GAT, or GATv2 layers. Unary and binary templates (e.g., $p(X)$, $p(X,Y)$) are scored via linear maps or DistMult-style bilinear forms; $$s_{unary}(p,X) = \sigma(\mathbf{a}_p^\top \mathbf{x}_X + b_p)$$ and $$s_{binary}(p,X,Y) = \sigma(\mathbf{x}_X^\top M_p \mathbf{x}_Y)$$. Training again uses binary cross-entropy loss.

A hybrid fallback approach employs the GNN for seen templates and reverts to the NLI model otherwise, yielding state-of-the-art F1 (81.0%) by combining GCN(UT+BT) and Llama2-13B (Li et al., 2024).

BERTSubs and OntoLAMA, as implemented in DeepOnto, operationalize these strategies for OWL ontologies, providing both fine-tuned classification and prompt-based (cloze/probing) inference frameworks for expansion (He et al., 2023).

Model/Method	Expansion Principle	Strengths
NLI + Verbalization	Supervised, semantic entailment	World knowledge, generics
GNN + Concept Embedding	Graph-based template instantiation	Pattern mining, domain terms
Prompt-based (OntoLAMA)	LM probing via verbalized prompts	Zero/few-shot, mask-filling
Hybrid (NLI+GNN)	Fallback composition	Best-of-both, adaptive

2. Statistical and Semantic Pattern-Based Enrichment

An alternative paradigm leverages statistical and linguistic regularities in large corpora to discover missing ontology terms and relations:

• Corpus-Driven Candidate Discovery: Raw web text is processed using NER, n-gram tokenization, and matching against the ontology to identify out-of-vocabulary n-grams as candidates.
• Statistical Relatedness (NTR): For each candidate, normalized term relatedness is computed using web hit counts:

$$X \equiv \text{Biologist} \sqcap \exists \text{livesIn}.\text{UK}$$0

where $$X \equiv \text{Biologist} \sqcap \exists \text{livesIn}.\text{UK}$$1 is the hit count for $$X \equiv \text{Biologist} \sqcap \exists \text{livesIn}.\text{UK}$$2, $$X \equiv \text{Biologist} \sqcap \exists \text{livesIn}.\text{UK}$$3 for co-occurrence, $$X \equiv \text{Biologist} \sqcap \exists \text{livesIn}.\text{UK}$$4 the indexed corpus size.

• Lexico-Syntactic Pattern Mining: Candidate-sense pairs with high NTR undergo pattern queries (e.g., "X is a Y", "X is part of Y"). The best-supported pattern determines the relation label (hypernym, hyponym, etc.). If all patterns fail, a generic "related to" link is inserted.
• Integration: For polysemous senses, attachments are disambiguated against sense subtrees using NTR or context overlap.

Empirical evaluation yields high precision (69–84%) across domains, though the recall is not measured (Maree et al., 2020). Limitation arises from reliance on a fixed pattern set and web search instability.

3. Modular, Profile-Driven, and Pattern-Oriented Expansion

Expansion by modular extension, profile definition, and leveraging external vocabularies is central in application domains such as cyber-physical systems:

• Gap Analysis & Selective Import: Identify deficiencies in core ontologies by modeling domain-specific scenarios. For example, the SCOPE paradigm extends UCO and CASE with Smart City Infrastructure concepts via OWL subclassing and equivalent class mappings.
• Integration of External Vocabularies: Import MITRE ATT&CK, CAPEC, and ISO standards as modular profiles. Local subclasses/individuals are wrapped around external IRIs to ensure interoperability.
• Pattern-Driven Axiom Design: Standard OWL patterns such as n-ary relations for evidence linkage, property restrictions (e.g., "hasThreatTechnique some mitre:Technique"), and annotation patterns for external IDs facilitate systematic enrichment.
• Validation: Use scenario modeling, competency questions (expressed as SPARQL queries), and comparative RDF serializations across baseline and expanded ontologies to demonstrate utility and expressivity, as showcased in smart city forensic investigations (Tok et al., 2024).

Best practices include modularization, non-destructive subclassing, close alignment with external standards, and scenario-driven evaluation.

4. Algebraic and Category-Theoretic Merging

Ontology expansion through algebraic closure formalizes the systematic combination of ontology repositories:

• Ontology Merging System $$X \equiv \text{Biologist} \sqcap \exists \text{livesIn}.\text{UK}$$5: A set of ontologies $$X \equiv \text{Biologist} \sqcap \exists \text{livesIn}.\text{UK}$$6, a binary alignment relation $$X \equiv \text{Biologist} \sqcap \exists \text{livesIn}.\text{UK}$$7, and a partial merge operator $$X \equiv \text{Biologist} \sqcap \exists \text{livesIn}.\text{UK}$$8 defined on aligned pairs.
• Algebraic Properties: Requirements:
  • Idempotence (I): $$X \equiv \text{Biologist} \sqcap \exists \text{livesIn}.\text{UK}$$9 and $$[CLS]~\text{verbalize}(X)~[SEP]~\text{verbalize}(Y)~[SEP]$$0
  • Commutativity (C): $$[CLS]~\text{verbalize}(X)~[SEP]~\text{verbalize}(Y)~[SEP]$$1
  • Associativity (A)
  • Representativity (R)
• Closure Algorithm: For a finite seed set, compute its merging closure $$[CLS]~\text{verbalize}(X)~[SEP]~\text{verbalize}(Y)~[SEP]$$2 via ascending chains, yielding a finite poset under the merging order. Maximal elements correspond to fully integrated ontologies, while minimal elements serve as atomic units.
• Instantiation via Pushouts: Formalizes merging as the categorical pushout over alignment cospans. Theoretical results guarantee finiteness, efficient computation, and algorithmic support for sorting and querying within the closure (Guo et al., 2022).

This algebraic perspective is applicable in domains (e.g., geospatial ontologies) where systematic integration and analysis of all possible merged configurations are required.

5. Ontology Expansion in Conversational Understanding

In dialogue systems and conversational AI, ontology expansion (OnExp) includes:

• New Intent Discovery (NID): Identifies both known and novel intents from user utterances. Techniques involve clustering (K-Means, DEC, DCN), contrastive learning (SCCL, DPN, RAP), LLM-based (in-context prompts, ChatGPT grouping), and hybrid few-shot methods. Evaluation employs ACC, ARI, and NMI on benchmarks such as BANKING77, CLINC150, and StackOverflow. ALUP achieves ACC/ARI/NMI of 82.9/73.1/88.4 (Liang et al., 2024).
• New Slot-Value Discovery (NSVD): Extracts novel slot types and values via unsupervised frame-semantic parsers, iterative clustering, and prompt-based methods. Architectures include sequence taggers, span-pointer networks, and contrastive prototype matchers. Partially supervised methods (e.g., GZPL) obtain Span-F1 up to 61.1 on SNIPS.
• Joint OnExp: Co-discovery of intents, slots, and values is addressed via coarse-to-fine multistage pipelines (e.g., RCAP) but remains challenging due to error propagation and unified training complexity.

Leading future directions are early-stage/few-shot OnExp, multimodal integration, holistic end-to-end systems, and LLM-based prompting (Liang et al., 2024).

6. Implementation Tools and Best Practices

Tooling support for ontology expansion is illustrated by DeepOnto, which provides:

• Transformer-based Subsumption (BERTSubs): Fine-tuning on (verbalized) subclass pairs within an ontology using a two-layer MLP head, negative sampling, and threshold-based candidate acceptance.
• Prompt-based Probing (OntoLAMA): Cloze-style prompts (e.g., "{C} is a kind of {D}. [MASK].") posed to masked LMs; zero/few-shot scoring and insertion of high-confidence predictions as axioms.
• Core Components: OntologyVerbaliser for recursive EL expression verbalization, OntologyNormaliser for axiom normalization, and robust negative sampling with reasoner-aided filtering.
• Extensibility: Users can swap in any Huggingface model, alter loss functions, or add GNN heads (He et al., 2023).

Key workflow recommendations are:

• Perform gap analysis on core ontologies through realistic modeling and user studies.
• Extend ontologies by subclassing/equivalence, avoiding direct modification.
• Modularize extensions for selective usage.
• Reuse and wrap established external vocabularies via owl:imports and subclassing.
• Evaluate via scenario-driven competency questions and SPARQL querying (Tok et al., 2024).

7. Limitations and Prospective Research Directions

Current limitations include:

• For GNN/template-based methods, candidate space is template-bound; generative approaches for rule induction are needed (Li et al., 2024).
• Integration strategies for combining statistical and semantic views typically use hard cutoffs rather than learned weighting, motivating attention-based or linear fusion methods (Li et al., 2024).
• Statistical/pattern-based methods have limited recall and sensitivity to web dynamics (Maree et al., 2020).
• Joint ontology expansion (e.g., in conversation) has challenges in error propagation, knowledge sharing, and unified benchmarking (Liang et al., 2024).
• Ongoing research focuses on few-shot adaptation, multi-modal ontological enrichment, and evaluation via downstream task success metrics.

Overall, ontology expansion leverages hybrid methodologies—deep inference, statistical analysis, modular pattern extension, and algebraic merging—to systematically enhance and adapt knowledge representations across domains. Combining complementary paradigms, scenario-driven evaluation, and open-ended candidate generation represents the current frontier for automated, robust ontology enrichment.