Metadata and Vocabulary for Knowledge Representation Learning

Knowledge Representation (KR) is foundational to AI, enabling machines to encode, store, and reason about information in ways that support complex tasks like reasoning, decision-making, and problem-solving [3], [18]. KR consists of countless methods and tools, from logical formalisms to ontologies, that structure knowledge for intelligent systems design and operation. One of the important KR artifacts are vocabularies (Figure 1), which standardize the choice of terms and sometimes their formal definitions. They are used to declare explicit existential descriptors of components and data used by automated intelligent systems and processes, and one of their main functions in resource description as in metadata. Explicit, shared and standardized Knowledge Representation in AI is a mechanism for creating human-readable, explainable, and auditable data models [5]. Despite its centrality, explicit and shared KR is rarely used to design modern AI systems, particularly machine learning (ML). This omission can lead to increased risks as the lack of resource description limits the transparency and auditability needed for trustworthy AI [19].

KR methods have evolved from symbolic logic and expert systems to ontologies and knowledge graphs. Key approaches where vocabularies can be useful include:

• Logical Representation: Uses formal languages to encode facts and rules, enabling inference [3]. = vocabulary of formal language
• Semantic Networks: Graph-based models of concepts and relationships [18]. = vocabulary defining concepts and relationship
• Frames: Structures for organizing stereotypical situations or objects [5]. = vocabulary naming structures
• Ontologies: Formal specifications of domain concepts and relations, supporting interoperability [11]. = vocabularies as part of the ontology

KR languages include logic-based languages (e.g., Description Logics), frame-based representations, semantic networks, production rules, ontologies (e.g., OWL), conceptual graphs, and structured natural language [1, 20]. All of these languages use vocabularies, specially controlled vocabularies and metadata to codify the use of terms.

KR is essential for enabling AI to reason and interact with the world, underpinning advances in expert systems, the Semantic Web, and modern AI applications [13]. However, a disconnect remains between traditional KR and its application in ML-driven AI.

Explicit KR is crucial for mitigating AI risks by encoding knowledge in human-readable formats, enabling audit, analysis, and evaluation by diverse stakeholders [8]. It provides a “source of truth” accessible to developers, domain experts, and end-users, making AI outcomes more interpretable. Integrating KR with ML can enhance risk identification and mitigation by combining expert knowledge with data-driven analytics [12]. Structured KR also supports governance and compliance by documenting policies, ethics, and regulatory requirements.

Developing a KR vocabulary involves domain analysis, modeling, and ontology engineering. Semantic modeling platforms leverage NLP, formal concept analysis, and entity extraction to create robust vocabularies that support reasoning and interoperability [1]. Creating domain-specific vocabularies and metadata sets requires collaboration with experts, analysis of domain needs, and computational knowledge extraction [16]. A core set of terms and definitions that covers the entire AI KR domain in all its facets is going to result of ongoing refinement and scoping based from knowledge of facets and co-occurring terms.

Metadata sets—also known as metadata schemas, element sets, or standards—play a critical role in knowledge representation by providing a structured way to describe, classify, and relate data. Their primary function is to enable semantic clarity, interoperability, and discoverability across systems, particularly in digital information environments, linked data, and the semantic web. Metadata sets define a consistent model for how knowledge about resources is described and organized, usually using elements (e.g., title, creator, subject, date). “Metadata provides the structured description of information objects, making it possible to retrieve and reuse them effectively.” [10] They allow diverse systems to understand and interpret data the same way by using a shared vocabulary and structure. “A key function of metadata is to support semantic interoperability among heterogeneous systems.” [15]

When metadata sets are aligned with ontologies (e.g., in RDF/OWL), they support automated reasoning, enabling machines to infer new knowledge from existing metadata. Metadata sets help express relationships among resources, making them integral to intelligent reasoning on the web. A subset of the intended outcome (vocabulary for KRL together with a proposed approach for its evaluation) is presented in this paper, however it does not, nor it is intended to be exhaustive of all possible outcomes.

Metadata uses terms from vocabularies to express knowledge within a formal knowledge representation framework.

As such, vocabularies that describe AI-based inputs, processes and outcomes, can be used to support the development of metadata, as well as providing standard mechanisms to describe the resources associated with AI.

During a routine search for terminology in the current standards for AI database,¹ a key concept from the KR domain such as \"truth preservation\" [9] is missing, not to be found in any of the AI standards in development. These focus on technical definitions, risk management, and compliance but omit critical system safety functions. Influential standards (ISO/IEC JTC 1/SC 42, IEEE, NIST, CEN/CENELEC) lack sufficient KR provisions to minimize AI risks. Further review confirms that concepts like \"truth maintenance\" are not formalized in current standards, despite their foundational role in classical AI [9, 6]. In this paper this omission is flagged as potentially leading to critical AI systems failures and sets out to search for other critical gaps.

In ML, belief propagation and probabilistic inference replace classical truth maintenance, but do not provide explicit consistency management or traceability [17, 18]. The lack of TMS-like mechanisms leads to:

• Inconsistent knowledge management
• Reduced explainability
• Difficulty handling contradictions
• Limited dynamic updates

These limitations need to be understood and addressed to mitigate AI risks. Addressing these shortfalls and gaps requires importing concepts from reliable knowledge systems and reliability engineering which is work in hand at the W3C AI KR CG [7].

An initial flat list of KR terms, aligned with the W3C AI Knowledge Representation Community Group’s goals, is being compiled and is being discussed elsewhere. The creation of unique definitions for each term is remanded at a further stage.

The vocabulary aims to first identify core concepts that are required to represent knowledge in machine learning, and will be used as a benchmark to identify gaps between KR and other domains (such as for example AI safety). Eventually a set of unique definitions is going to be associated with each term and content, addressing at least in part the semantic gap in AI standards. The first subset of the vocabulary presented here is scoped through domain facet identification and use case analysis. It is not intended to cover application-specific or domain-specific terms but to capture some aspects of foundational KR concepts essential to query, explain and evaluate machine learning processes. This is deemed essential to attempt auditable trustworthy and human-readable assessments necessary to identify AI risks. Creating a domain vocabulary for knowledge representation and a metadata set tailored to specific KR tasks and targeted subdomains involves systematic processes to ensure representativeness, interoperability, and support for data management, discovery, and reuse. Common methods adopted in the development of the vocabulary under development include:

Domain-Specific Metadata Scheme Development: Collaborate with domain experts to design a metadata scheme that captures all necessary concepts without overburdening users [16].

Domain Analysis and Metadata Element Identification: Study the domain to define the scope, resource types, and user needs, identifying relevant metadata elements [16].

Semantic Knowledge Extraction and Knowledge Discovery: Use computational methods to extract and formalize domain knowledge [13].

KR is vast and interdisciplinary, with overlapping terminology across mathematics, logic, computer science, and engineering [3]. The lack of a standardized KR lexicon complicates communication and hinders explainability, transparency, and trustworthiness in AI. A major challenge is defining the core body of knowledge and identifying essential co-occurring terms within KR’s many facets, as well as to handle the vast and heterogeneous datasets (Figure 4). This paper discusses the challenge by scoping the work tightly and by limiting focus on presenting a subset of terms for evaluation and discussion (See Appendices).

This vocabulary development presented here is carried out by using a combination of methods:

• Domain Scoping: Defining the KR body of knowledge and selecting authoritative sources [3]
• Facet Identification: Analyzing domain facets and use cases to select representative terms.
• Expert Collaboration: Engaging with domain experts and standards organizations.

Future iterations will refine the vocabulary subset, populate a master list of global terms that capture the extended AI KR domain and will provide formal definitions that can be used in learning for the purpose of disambiguation, teaching and professional development.

A wide and disparate range of knowledge sources from AI KR, from classic textbooks which laid the foundations of this foundational epistemic field of knowledge from its beginnings, to the most contemporary uses in machine learning, is an extensive list of terms, numbering in the thousands.

The terms, once compiled in a long masterlist, cover most of the AI KR in use in literature. However there are variations in definitions of these terms in each knowledge source, and each term can point to concepts that can have either a very generic meaning applicable to diverse disciplines, including say, the broader aspects of computer science and information systems, to very specific interpretations and uses that can be narrowed down to answering the explainability questions:

What data and knowledge are used by the algorithm? In what way? Following what processes? With what outcomes?

This work focuses on the latter goal, to list terms and concepts that can help to answer the explainability questions. To facilitate the human readability aspect and cognification of a long flat lists of terms, a first layer of abstraction aims to identify categories and metatags that can be used to categorize the terms, as in Figure 5.

Secondly, each facet is used to shape a subset of the terms in use (remember that the master list is very long), that is a set of separate vocabularies of terms that although have common uses in multiple domains and facets, are presented to characterize a specific context.

For example, as in the case presented in Appendix A, a vocabulary representing the domain KRL (Knowledge Representation Learning). Knowledge Representation Learning bridges symbolic AI (structured knowledge) and statistical AI (machine learning) by embedding symbolic structures into vectors. It enables more effective reasoning, inference, and generalization across a wide range of AI applications.

In our work, KRL vocabulary lists terms in use in ML tagged with corresponding concepts in symbolic AI (Appendices B, C).

A number of use cases extracted from literature are being adopted to verify the validity and usefulness of the outcomes in terms of adequacy of coverage, with these specific questions in mind:

In attempting to answer the evaluation questions for each use case (what does the AI system do, what data/knowledge does it use, which processes does it follow): Are there key terms in the use cases that are missing from the vocabulary? Can the candidate terms in the proposed draft vocabulary adequately describe what the AI system does and how?

Is the coverage of the terms identified adequate to represent the knowledge representation in the AI (noting that domain knowledge is not in scope of the vocabularies being developed)?

Sample Use Cases used for the evaluation include:

1. Medical Expert Systems: Low Back Pain Management (Santra et al., 2018)
2. Manufacturing Decision Support: Actionable Cognitive Twins (Rozanec et al., 2021)
3. Intelligent Personal Assistants: CALO Project (Cheyer, 2005)
4. Ontology-Based Explanation in Clinical AI (Chari et al., 2020)
5. Biomedical knowledge analysis and recommender system
6. Legal decision support

The usefulness of standardizing the terminologies of knowledge representation in AI, to support the explicit and shared description of associated resources and processes, thus facilitating their explainability and trustworthiness, is consistently proven and part and parcel of information technology. The preliminary outcome of the KRL vocabulary definition confirm the importance and role of such practices.

Explicit, standardized KR is essential for explainable, trustworthy, and auditable AI. Current AI standards lack critical KR concepts, increasing system risks. This paper’s initial vocabulary is an example of work being undertaken toward bridging the semantic gap and supporting the development of reliable AI systems.

Future work includes identifying further terms (from ongoing evaluations), attempting to produce unique meaningful definitions, and derive further levels of abstraction as meta tags pointing to higher order categories of logic and process definition.

See also Appendices A, B, C.

• architecture reformulation
• bag-of-words
• begin-inside-outside
• bidirectional
• bidirectional long short-term memory
• big model systems for large-scale representation learning
• chain-of-thought
• complexity of an algorithm
• compositional semantics
• computer vision
• concatenate a list of vectors/tensors
• conditional random field
• continuous bag-of-words
• convolution operator
• convolutional neural network
• cosine similarity
• cross-lingual representation learning
• cross-modal representation
• deep convolutional generative adversarial networks
• deep high-resolution representation learning
• distillation
• domain adaptation
• domain transfer
• element-wise multiplication
• encoder representations
• expectation-maximization algorithm
• exponential function
• feed-forward network in transformers
• feed-forward neural network
• gated recurrent unit
• generative pre-trained model
• gradient of a function
• graph convolutional network
• graph processing unit
• graph representation learning
• hidden markov model
• hidden representation
• high-to-low resolution subnetworks
• hyperbolic tangent function
• information extraction
• information maximizing generative adversarial nets
• input augmentation
• interpretable representation learning
• inverse document frequency
• knowledge base
• knowledge graph
• knowledge-guided NLP
• kronecker product of two matrices
• language model
• latent dirichlet allocation
• layer norm layer in transformers
• limitation
• logical
• loss function
• LSTM long short-term memory
• machine translation
• mask language modeling
• mixture-of-experts
• momentum contrast for unsupervised visual representation learning
• momentum contrastive learning
• multi-layer perceptron
• multi-resolution subnetworks in parallel
• multilayer perceptron
• named entity recognition
• natural language inference
• natural language processing
• neural discrete representation learning
• neural probabilistic language model
• norm of a vector
• normal distribution
• objective regularization
• ordinary differential equations
• parallel distributed processing
• parameter transfer
• part-of-speech
• partial derivative of a function
• pointwise mutual information
• pre-trained language model
• pre-trained model
• proportional
• pruning
• quantization
• rectified linear unit activation function
• recurrent neural network
• relation extraction
• representation learning
• retrosynthetic prediction
• robust representation
• self-attention layer
• sememe-based lexical knowledge representation
• sequence tagging
• sigmoid function
• similarity
• single-model networks
• singular value decomposition
• size of a set
• softmax function
• support vector machine
• tagging
• term frequency-inverse document frequency
• transformers
• transpose
• unsupervised representation learning
• visual question answering

Formal Logic: Model, Proof, QBF, SAT

KR: Default Logic, Frame, Model, Schema

Logical: Clause, Contradiction, Ground, Inference, Negation, Nonmonotonic, Predicate, Skolemization, Variable

Nonmonotonic: Default Logic

Person: Reiter

Programming: Function

Reasoning: Backtracking, Inference, Nonmonotonic, Resolution, Suppression Task

Structure: Frame, Schema

Syntax: Clause, Function, Ground, P(c,x) (predicate example), Predicate, Skolemization, Variable