Comprehensive Landscape Study and Synthesis for Embedded Semantics

Full report

On this page

1. Executive Summary

The transition from lexical data structures to semantic information systems has been historically fragmented by competing architectural paradigms. Symbolic systems, encompassing knowledge graphs and formal ontologies, offer high precision, stable identities, and rigorous logical structures, yet they consistently suffer from systemic brittleness, exorbitant curation costs, and a fundamental inability to scale across the fluid nuances of natural language. Conversely, connectionist systems, defined by dense vector embeddings and large language models (LLMs), deliver remarkable zero-shot generalization and robust semantic retrieval, but they fail categorically at maintaining stable conceptual identity, tracking provenance, versioning knowledge, and expressing explicit ambiguity. Dense vectors, by their mathematical nature, conflate distinct concepts into overlapping lower-dimensional spaces, resulting in semantic residue—the nuanced stylistic, contextual, or peripheral meaning that is lost or entangled during projection1. Embedded Semantics represents a novel architectural synthesis designed to resolve this dichotomy by explicitly decoupling semantic identity from latent vector representations. By defining vectors purely as volatile evidence rather than the immutable unit of identity, the Embedded Semantics framework establishes a stable, registry-backed foundation for meaning representation. This landscape study exhaustively analyzes adjacent systems—ranging from classical lexical databases such as WordNet and FrameNet3 to translation interlinguas like the Universal Networking Language (UNL)5, massive knowledge graphs including Wikidata and BabelNet6, biomedical terminologies such as the Unified Medical Language System (UMLS)8, and contemporary vector infrastructures. The analysis reveals that while isolated components of the Embedded Semantics philosophy exist in disparate domains, no existing system successfully unifies multilingual semantic retrieval, stable registry-backed Concept IDs, explicit abstention, positive and negative evidence constraints, and model-independent identity into a single cohesive framework. The resulting synthesis precisely positions Embedded Semantics not as a competitor to vector databases or LLMs, but as the essential semantic infrastructure required to render latent artificial intelligence representations stable, versionable, and cryptographically accountable over time.

2. Embedded Semantics in One Paragraph

Embedded Semantics is an architectural framework and meaning representation paradigm that fundamentally decouples semantic identity from latent vector representations. It assigns stable, registry-backed Concept IDs to meanings, treating the vectors derived from machine learning models strictly as volatile evidence of a concept rather than the concept itself. The framework uniquely integrates multilingual semantic retrieval, versioned meaning, explicit ambiguity, and the capacity for system abstention, ensuring that semantic meaning remains model-independent and rigorously auditable. By explicitly capturing semantic residue and supporting both positive and negative evidence constraints, Embedded Semantics bridges the high precision and provenance of symbolic knowledge graphs with the generalized retrieval power of connectionist artificial intelligence, creating a durable, interoperable semantic infrastructure for the next generation of intelligent systems.

3. Embedded Semantics in One Page

The foundational crisis in modern artificial intelligence is the persistent conflation of mathematical representation with semantic identity. In contemporary vector databases and embedding architectures, a concept is defined entirely by its coordinates in a high-dimensional continuous space. However, this geometric space shifts dramatically every time a model is retrained, updated, or replaced, effectively destroying the continuity of meaning across the system. Furthermore, dense vectors inherently discard nuanced meaning—often termed "semantic residue"—and fail to distinguish between highly related but definitionally distinct concepts, rendering these systems incapable of modeling explicit ambiguity or engaging in calculated abstention1. Embedded Semantics resolves this structural failure by anchoring meaning in stable, registry-backed Concept IDs. Within this paradigm, vectors are demoted from their assumed role as authoritative identities to serve merely as computational evidence. A specific dense embedding is recognized as just one model’s temporal, probabilistic observation of a Concept ID. This architectural separation unlocks several critical capabilities that are currently absent in the broader artificial intelligence ecosystem. First, the framework enables absolute model-independent semantic identity. A Concept ID remains constant regardless of whether the processing is executed by a local dense retriever, a massive autoregressive large language model, or a deterministic symbolic rules engine. Second, it introduces rigorous versioned meaning and provenance. Because conceptual identity is separated from transient vectors, the evolution of a concept’s definition, its compositional relationships, and its historical contexts can be versioned and tracked via centralized or decentralized registries, answering the critical questions of how and why a system arrived at a specific semantic conclusion. Third, Embedded Semantics natively supports explicit ambiguity and abstention. When a traditional vector system encounters a polysemous term like "bank," the resulting embedding often collapses the financial and geographical meanings into an averaged vector, or it arbitrarily commits to one interpretation based on statistical priors. Embedded Semantics allows an architecture to explicitly state that an input represents multiple conflicting Concept IDs, while acknowledging a lack of sufficient evidence to resolve the ambiguity. It further permits the system to abstain entirely, recognizing an entity without forcing a mapping to an unverified Concept ID11. Fourth, the architecture leverages compositional relationships and multilingual semantic retrieval. By intelligently combining discrete concepts, the system can represent highly complex states without requiring a monolithic, brittle embedding for the entire state. Multilingualism is treated as an inherent property because the Concept ID acts as an interlingua; the English, French, and Mandarin expressions of a concept map to the exact same immutable identifier, echoing the structural goals of BabelNet6 or UNL5, but achieving this through native integration into vector-based retrieval pipelines. Finally, by incorporating positive and negative evidence, Embedded Semantics empowers systems to define explicitly what a concept is not, effectively bounding the semantic space and preventing the hallucinations inherent in purely associative vector spaces. Embedded Semantics does not seek to replace vector search products or latent models; rather, it constitutes the semantic layer that governs them, providing the stability, provenance, and human-readable optional public renderings necessary for enterprise, legal, medical, and mission-critical applications.

4. Landscape Map

To position Embedded Semantics accurately within the broader ecosystem, it is necessary to map the surrounding technological terrain. The landscape fundamentally divides into three primary continents of meaning representation, each possessing distinct advantages and fatal structural limitations. The first continent comprises Symbolic and Lexical Systems, encompassing high-precision, low-recall architectures that rely on explicit human curation. Systems such as WordNet, the Semantic Web, FrameNet, and the Unified Medical Language System (UMLS) reside here. These frameworks possess highly stable identities, deep provenance, and formal logic, but they categorically fail at zero-shot semantic retrieval and struggle to scale across the fluid, analog boundaries of natural human discourse. The second continent consists of Connectionist and Latent Systems, featuring high-recall, low-precision architectures driven by statistical co-occurrence and deep learning. This includes vector databases, large language model latent spaces, and multilingual embedding architectures. These systems excel at fuzzy semantic retrieval and cross-domain generalization, but they fail entirely at maintaining stable identity, tracking versioned meaning, and expressing explicit abstention, treating volatile mathematical arrays as ground truth. The third continent represents the Interlingua and Bridging Archipelago, containing systems attempting to bridge human language and machine representation across linguistic barriers. Architectures like BabelNet, ConceptNet, the Universal Networking Language (UNL), and the Common Locale Data Repository (CLDR) operate in this space. They offer impressive multilingualism and structural compositionality, but they generally lack native vector-evidence integration and struggle to explicitly model the semantic residue lost during translation or abstraction. Embedded Semantics occupies the intersection of these three domains, serving as the definitive connective tissue. It adopts the strict, stable identity models of the Symbolic continent, applies them directly to the probabilistic retrieval mechanisms of the Connectionist continent, and utilizes the cross-lingual philosophies of the Interlingua archipelago to create a comprehensive, model-independent architecture.

5. WordNet

WordNet was designed to solve the problem of mapping the English lexicon into a machine-readable semantic network based on psycholinguistic principles, organizing words by meaning rather than morphology. Its fundamental unit of identity is the synset, a cognitive grouping of synonymous words representing a single discrete concept. While the original WordNet architecture lacks native multilingual labels, subsequent external extensions such as EuroWordNet and the Open Multilingual WordNet have attempted to address this deficiency. The system deeply supports semantic relationships, offering extensive mapping of hypernymy, hyponymy, meronymy, and antonymy. However, it offers minimal provenance, relying entirely on the central authority of its creators at Princeton without providing granular tracking for individual assertions. Versioning occurs only in static, discontinuous batches, and the system categorically fails to represent uncertainty, ambiguity, or abstention, relying instead on absolute lexical classifications. Furthermore, WordNet does not support native semantic retrieval, nor does it utilize vector representations, let alone treat them as authoritative. Its stable IDs, known as synset offsets, offer a clear point of overlap with Embedded Semantics, alongside the shared goal of capturing deep conceptual relationships. However, Embedded Semantics differs fundamentally by integrating continuous vector evidence, positive and negative constraints, and dynamic provenance. The crucial lesson Embedded Semantics must adopt from WordNet is the cognitive validity of grouping by meaning to form a foundational Concept ID. Conversely, the primary mistake to avoid is WordNet's rigid, brittle taxonomic hierarchy, which frequently collapses under the fluid, compositional reality of natural language and struggles to represent semantic residue.

6. Wikidata

Wikidata solves the problem of providing a collaboratively edited, multilingual, structured knowledge base to support Wikimedia projects and the broader open-data ecosystem, effectively acting as a central hub for global entity resolution. Its core unit of identity is the Q-node, a language-agnostic identifier assigned to every distinct entity. The system inherently supports multilingual labels, as a single Q-node possesses labels, aliases, and descriptions in hundreds of languages simultaneously. Wikidata extensively supports relationships through a rich, graph-based property system of P-nodes, linking entities into a massive web of knowledge. It excels in provenance, dictating that every statement should be accompanied by a reference pointing directly to a verifiable source. Versioning is rigorously maintained, with complete page histories and revision tracking available for every entity modification. Uncertainty is partially represented; the system supports "unknown value" or "no value" assertions and allows for deprecated ranks, which closely mirrors the concept of abstention. However, Wikidata does not natively support dense semantic retrieval; retrieval is executed via SPARQL, an exact-match logical query language, though third-party developers frequently build embeddings on top of its graph. Vector representations are entirely external and not considered authoritative by the system. Its use of stable, permanent IDs heavily overlaps with the philosophy of Embedded Semantics, as do its commitments to multilingualism, provenance, versioning, and structural abstention. Embedded Semantics differs by acting as a general semantic and linguistic meaning framework rather than an encyclopedic entity graph, actively ingesting vector evidence and explicitly modeling semantic residue. The Q-node architecture provides a masterclass lesson in language-agnostic, stable identifiers that Embedded Semantics should adopt. The primary mistake to avoid is the vulnerability of a massive crowdsourced ontology to contradictory modeling, necessitating stricter compositional logic and negative evidence constraints in Embedded Semantics.

7. BabelNet

BabelNet addresses the fragmentation of lexical and encyclopedic knowledge by seamlessly integrating WordNet, Wikipedia, Wikidata, and other resources into a massive, unified multilingual semantic network6. The foundational unit of identity within this architecture is the Babel synset, an abstract node representing a specific meaning across languages. It provides exhaustive support for multilingual labels, covering hundreds of languages by systematically merging human-curated interlanguage links with statistical machine translation6. BabelNet comprehensively supports relationships, directly importing and harmonizing relational data—such as WordNet lexical relations and Wikipedia structural hyperlinks—from its source networks12. Provenance is a core feature; the system meticulously tracks the origin of each translation and relation, differentiating between human-curated encyclopedic links and automatically generated semantic corpora6. Versioning exists, but has historically been a complex issue; managing the evolution of synsets across massive batch updates requires intricate version mapping infrastructures6. Uncertainty is partially modeled through the assignment of confidence scores to machine-translated relations, allowing users to filter based on evidentiary strength7. While researchers have generated vector embeddings based on its structure, BabelNet relies primarily on graph-based retrieval rather than native semantic vector retrieval, and it does not treat vector representations as authoritative. Its implementation of stable Babel synset IDs, multilingual meaning representation, and the integration of varying evidence confidence strongly overlaps with Embedded Semantics. However, BabelNet functions as a static aggregation of existing databases and does not actively manage semantic residue or define a real-time vector-evidence architecture. The integration of encyclopedic entities and lexicographic concepts is a vital lesson for Embedded Semantics to adopt. The critical mistake to avoid is relying too heavily on automated, unverified machine translation for mapping concepts, which can pollute the semantic registry with false positive evidence and degrade overall system precision.

8. ConceptNet

ConceptNet was engineered to solve the challenge of providing artificial intelligence systems with explicit common-sense knowledge—such as the understanding that water is used for drinking—to improve natural language understanding and logical inference. Its unit of identity is the node, representing normalized words or short phrases, connected by typed semantic edges. It actively supports multilingual labels, mapping concepts across multiple languages into a unified node representation. The system relies heavily on explicit relationships, utilizing a specific, curated set of common-sense relations including UsedFor, IsA, and CapableOf. ConceptNet provides robust provenance, recording whether an assertion originated from crowdsourced platforms, expert systems, or automated text extraction routines. Versioning is maintained at the broader dataset release level, though not dynamically per concept. Uncertainty is represented structurally via weights; each edge possesses a numerical weight indicating the reliability, frequency, or universality of the assertion. ConceptNet bridges closer to native semantic retrieval than traditional lexical databases by explicitly supporting and distributing ConceptNet Numberbatch embeddings, though the vector representation itself remains subordinate to the graph and is not the sole authoritative identity. Stable IDs are maintained as stable URIs based on the normalized text of the concept. The system overlaps significantly with Embedded Semantics through its support for positive and negative evidence (utilizing relations like NotDesires and NotCapableOf) and its commitment to multilingualism. However, Embedded Semantics differs crucially in its architecture: ConceptNet's unit of identity is fundamentally tied to a normalized language-specific string, rendering it highly susceptible to polysemy and ambiguity. Embedded Semantics must adopt ConceptNet's powerful use of explicit negative evidence to bound AI behavior, while studiously avoiding the mistake of tying the fundamental unit of conceptual identity to a morphological string, which re-introduces the very conflation of representation and meaning that Embedded Semantics seeks to eliminate.

9. FrameNet

FrameNet implements the linguistic theory of Frame Semantics, solving the problem of defining word meanings based on the cognitive "frames"—situations, events, or scenarios—they evoke, and explicitly mapping the semantic roles involved in those scenarios3. The unit of identity is split between the abstract Semantic Frame and the specific Lexical Unit that evokes it. While the core project is fundamentally English-centric, independent international efforts have successfully mapped other languages to these same abstract frames, demonstrating implicit support for multilingual labels4. FrameNet utilizes a highly sophisticated network of frame-to-frame relationships, capturing Inheritance, Subframes, Causative relations, and Perspectives3. Provenance is a cornerstone of the architecture; the database is rigorously corpus-based, providing specific annotated sentences as empirical proof of a frame's existence and structural valence16. Versioning occurs at the static release level, and the system lacks any formal modeling of uncertainty, relying on discrete, deterministic human annotations. FrameNet does not support native semantic vector retrieval, nor does it acknowledge vector representations as authoritative. It does, however, maintain stable identifiers for both Frames and Lexical Units. FrameNet's architectural separation of core frame elements from non-core, peripheral elements provides a strong conceptual overlap with Embedded Semantics' approach to tracking semantic residue—the specific, contextual details of a generalized event17. The systems differ in that FrameNet focuses heavily on predicate-argument structures and verbs, whereas Embedded Semantics provides a universal architecture for all semantic concepts, driven by latent vector evidence. The vital lesson for Embedded Semantics is that meaning is deeply compositional and contextual, often requiring a "frame" of surrounding entities to be fully resolved. The primary mistake to avoid is the extreme granularity in role definitions that plagues FrameNet, which frequently leads to sparse data distribution and insurmountable manual annotation requirements.

10. AMR

Abstract Meaning Representation (AMR) solves the problem of syntactic variance by providing a semantic representation language that captures the meaning of a whole sentence as a rooted, directed, acyclic graph, deliberately abstracting away from morphological and syntactic idiosyncrasies. Its fundamental units of identity are abstract concepts—frequently based on PropBank rolesets—and their connecting semantic roles. While AMR was designed with the intention of serving as an interlingua, it remains heavily biased toward English syntactic structures, making cross-lingual AMR an active but imperfect research area. AMR relies entirely on deep compositional relationships, generating exhaustive semantic networks within the boundaries of a single sentence graph. Provenance is implicitly tied to the specific source text corpus being annotated, and versioning is handled strictly at the treebank dataset level. AMR completely rejects uncertainty; the annotation guidelines force human annotators to commit to a single interpretation of a sentence, actively stripping away ambiguity. It does not support native semantic vector retrieval, nor does it treat vector representations as authoritative evidence. AMR does utilize stable IDs for its underlying concepts via PropBank identifiers. The system overlaps with Embedded Semantics in its fierce commitment to isolating pure meaning from syntactic structure and its reliance on complex compositional relationships. However, AMR is constrained to modeling sentence-level propositional logic, whereas Embedded Semantics models the fundamental architecture of isolated concepts and their continuous latent evidence across entire systems. The abstraction of "who did what to whom" into a language-agnostic graph is a vital lesson that Embedded Semantics should adopt. Conversely, AMR’s primary flaw—forcing annotators or systems to artificially resolve ambiguity even when they lack sufficient textual evidence—is exactly the mistake Embedded Semantics avoids by embracing explicit ambiguity and systemic abstention.

11. Semantic Web

The Semantic Web, encompassing standards such as RDF, OWL, SKOS, and schema.org, was designed to solve the fragmentation of internet data by providing a common framework that allows data to be shared, integrated, and reasoned over across application, enterprise, and community boundaries, rendering the web machine-readable19. The absolute unit of identity across the Semantic Web is the Uniform Resource Identifier (URI). It offers robust support for multilingual labels, as string literals attached to URIs can be explicitly tagged with language codes. Relationships are the foundational building blocks of the system, constructed entirely of RDF triples that connect subjects, predicates, and objects into a vast graph20. Provenance is highly supported through mechanisms like reification, RDF-star, and named graphs, which allow for detailed metadata tracking of individual assertions. Versioning is handled via specific tracking ontologies or dataset-level versioning, though it can be cumbersome. Uncertainty is notoriously difficult to model natively in this ecosystem; while probabilistic extensions to OWL exist in research, the core Semantic Web relies on strict, deterministic logic. It completely lacks native semantic vector retrieval, utilizing SPARQL for exact-match logical querying, and it categorically rejects vector representations as authoritative, relying entirely on the URI and the triple. The Semantic Web deeply overlaps with Embedded Semantics in its pursuit of model-independent semantic identity, the utilization of stable registry-backed IDs, and the reliance on compositional relationships. However, the Semantic Web completely ignores the reality of dense vectors, assuming a purely symbolic world. Embedded Semantics acknowledges that vectors are the dominant currency of modern artificial intelligence and provides a framework to manage them as evidence. Embedded Semantics must adopt the Semantic Web's use of URIs as a mechanism for global uniqueness and resolution. The critical mistake to avoid is the Semantic Web's requirement for rigid formal proofs (OWL) for every assertion, which historically prevented mass adoption; Embedded Semantics must remain lightweight, focusing on probabilistic vector evidence rather than strict logical completeness.

12. Medical Terminologies

Medical terminologies such as the Unified Medical Language System (UMLS) and SNOMED CT solve the critical problem of clinical interoperability by providing standardized, highly controlled vocabularies for documentation, billing, and research, ensuring that differing medical terms map to the exact same underlying clinical concept8. The unit of identity is the Concept Unique Identifier (CUI) in UMLS9 and the Concept ID in SNOMED CT. Multilingual labels are heavily supported, with UMLS systematically integrating translated vocabularies from multiple global sources21. Relationships are extensively mapped; the UMLS Semantic Network provides overarching Semantic Types and relations, while SNOMED CT utilizes the Expression Constraint Language (ECL) for highly complex compositional queries8. Provenance is absolute; UMLS tracks exactly which source vocabulary provided a specific concept or relation. Versioning is exceptionally rigorous, including strict protocols for concept retirement, deprecation, and historical mapping. Uncertainty is generally limited to specific, pre-coordinated clinical findings rather than systemic architectural ambiguity. These systems do not support native vector-based semantic retrieval, though they are heavily utilized as targets for NLP entity linking, and they do not view vector representations as authoritative. The use of highly stable IDs is the defining feature of these systems. Medical terminologies share massive overlap with Embedded Semantics, particularly in their use of stable registry-backed Concept IDs, rigorous provenance, versioned meaning, compositionality, and the explicit handling of ambiguity and polysemy. They differ primarily in scope; they are domain-specific ontologies devoid of native vector integration, whereas Embedded Semantics is a domain-agnostic architecture encompassing vector evidence. The UMLS Metathesaurus model—where a CUI serves as an abstract anchor for multiple source strings and definitions—is precisely the architectural lesson Embedded Semantics must adopt. The mistake to avoid is the trap of SNOMED CT's post-coordinated expressions, which can become overly complex and lead to multiple disparate ways of representing the exact same concept, ultimately defeating the purpose of standardization.

13. Embedding Systems

Embedding systems, particularly Large Language Model (LLM) latent representations, solve the problem of zero-shot generalization and natural language understanding by mapping discrete text tokens into high-dimensional continuous spaces, capturing rich semantic, syntactic, and contextual relationships through statistical co-occurrence. The fundamental unit of identity in these systems is the dense vector—an array of floating-point numbers. They support multilingual labels implicitly; massive multilingual models organically map different languages into shared latent spaces based on training data. Relationships are not explicitly defined but are implicitly encoded and retrieved via cosine similarity and attention mechanisms. These systems possess absolutely no provenance; once a concept is embedded in the weights of an LLM, the source of that knowledge is permanently lost and unauditable. Versioning is highly destructive; retraining a model or updating weights entirely shifts the latent space, meaning vectors from one model version cannot be directly compared to vectors from a subsequent version. Uncertainty is represented mathematically via probability distributions and logits, though LLMs are notoriously prone to overconfident hallucinations. Semantic retrieval is supported natively and powerfully. Crucially, these systems treat the vector representation as entirely authoritative. Consequently, there are no stable IDs; identity is entirely fluid. Embedding systems overlap with Embedded Semantics solely in their capacity for rapid semantic retrieval and implicit multilingualism. However, they violate almost every structural principle of Embedded Semantics. They treat vectors as identity, completely destroy provenance, lack stable IDs, and suffer massively from semantic residue by conflating distinct stylistic and semantic features into a single coordinate2. The lesson Embedded Semantics must adopt is that continuous vector representations capture the fuzzy, analog nature of human language far better than any symbolic system ever has. The catastrophic mistake to avoid is allowing the mathematical embedding to become the conceptual identity; the vector must remain rigorously subordinated as evidence pointing to a stable ID.

14. Vector Databases

Vector databases and search products, such as Pinecone, Milvus, and Weaviate, solve the infrastructural problem of providing highly efficient, scalable environments for storing, indexing, and querying massive volumes of high-dimensional vectors, utilizing algorithms like Hierarchical Navigable Small World (HNSW) graphs. The unit of identity in these systems is the Vector itself, typically paired with a UUID payload. Multilingual support is entirely dependent on the specific model used to generate the embeddings prior to storage. These databases contain no native semantic relationships, operating solely on nearest-neighbor geometric relationships. Provenance and versioning are entirely payload-dependent, relying on external application logic, as the vector spaces themselves do not version meaning natively. Uncertainty is not explicitly modeled; geometric distance is merely used as a blunt proxy for relevance or confidence. Semantic retrieval is the explicit, native function of these systems. Crucially, vector databases treat the vector representation as highly authoritative; the database architecture fundamentally assumes the vector is the ground truth for search and retrieval. While UUIDs can be stable, they typically map to a specific document, chunk, or string, rather than an abstract semantic concept, meaning the IDs fluctuate if the embedding model changes. The primary overlap with Embedded Semantics is the foundational utilization of vectors for rapid, scalable retrieval. Embedded Semantics differs structurally by acting as the semantic governance layer directly above the vector database. Embedded Semantics dictates that the vector database stores temporal evidence of a Concept ID, not the identity itself. The vital lesson for Embedded Semantics is that Approximate Nearest Neighbor (ANN) search is mandatory for operational scale. The critical mistake to avoid is designing an application architecture that irreparably breaks and loses all semantic indexing when the underlying embedding model is inevitably upgraded.

15. Multilingual Models

Multilingual models, including architectures like mBERT, XLM-R, and Cohere Multilingual, solve the problem of cross-lingual natural language processing by projecting multiple languages into a single, shared embedding space, allowing a model trained predominantly on English data to perform zero-shot inference on languages like Swahili or Hindi. The unit of identity is the multilingual dense vector. They inherently support multilingual labels by design. Semantic relationships are implicit, calculated via vector mathematics in the shared space. Like all latent systems, they lack provenance and native versioning capabilities. Uncertainty is measured via standard softmax probabilities. Semantic retrieval is highly supported, and the vector representation is treated as the authoritative ground truth. They do not utilize stable IDs. Multilingual models overlap with Embedded Semantics in their goal of achieving seamless multilingual semantic retrieval. However, Embedded Semantics differs fundamentally in execution: while it codifies cross-lingual meaning into a stable, auditable Concept ID, multilingual models merely place translated text in geometric proximity within a latent space, leaving them highly subject to semantic drift and the entanglement of semantic residue. The lesson Embedded Semantics must adopt is that shared latent spaces are exceptionally powerful and efficient evidence generators. The mistake to avoid is the assumption that because two translations are geographically close in vector space, they denote the exact same conceptual meaning. Cultural nuances and semantic residue often distinguish them, requiring the explicit, discrete modeling that only Embedded Semantics provides.

16. Translation Interlinguas

Translation interlinguas, exemplified by the Universal Networking Language (UNL), solve the problem of machine translation by providing a declarative formal language meant to represent core semantic data extracted from natural language texts, acting as a language-agnostic pivot point5. The unit of identity is the Universal Word (UW). UNL supports multilingual labels robustly; UWs are expressed using English root words appended with specific constraint lists for disambiguation, but they represent universal concepts mapped algorithmically to target languages5. The system deeply supports relationships, using binary semantic relations to construct comprehensive hypergraphs of meaning25. Provenance is implicitly tied to the specific source text being analyzed, while versioning occurs through batch dictionary updates rather than dynamic adjustments. UNL limits its handling of uncertainty by explicitly focusing on literal, direct meaning while systematically abstaining from processing poetry, metaphor, and subtle innuendo—an approach that handles semantic residue through intentional exclusion25. Semantic retrieval is possible by matching UNL graphs to measure semantic textual similarity, actively disregarding syntactic differences26. Vector representations are not utilized and are therefore not authoritative. Stable IDs are maintained via the Universal Words. UNL overlaps significantly with Embedded Semantics in its pursuit of model-independent meaning, the use of stable IDs, and its intentional limitation of scope (where UNL drops nuance, Embedded Semantics explicitly models it as semantic residue). The primary difference is that UNL relies on explicit parsing rules and syntactic dependency mapping rather than machine learning embeddings5. Using English as a human-readable base for an ID, augmented with constraints, is a pragmatic lesson Embedded Semantics should adopt, provided the ID remains structurally abstract. The mistake to avoid is relying entirely on deterministic, rule-based parsers to generate semantic graphs, which scale poorly; Embedded Semantics must leverage modern embeddings as the primary evidence generators.

17. Symbol Systems

Symbol systems, primarily the Unicode Standard and the Common Locale Data Repository (CLDR), solve the foundational problem of software internationalization by standardizing character encoding and providing standard building blocks for adapting software to diverse linguistic and cultural conventions27. The unit of identity is the Code Point in Unicode and the Locale ID in CLDR. These systems are inherently multilingual, defining semantic locale data identifiers across languages for elements like date formats, currency symbols, and pluralization rules29. Relationships are managed through rigid hierarchical fallback structures (e.g., en-US falling back to en)28. Provenance is absolute, maintained rigorously by the Unicode Consortium. Versioning is highly structured and semantic, ensuring absolute backward compatibility30. Uncertainty is explicitly managed; CLDR includes intelligent fallback mechanisms and explicitly marks unconfirmed data as "draft," allowing systems to degrade gracefully32. These systems do not support semantic vector retrieval, nor do they utilize vector representations. Code points and Locale IDs are the ultimate examples of stable, globally standardized IDs. Symbol systems overlap with Embedded Semantics in their reliance on global standard registries, rigorous versioning, and the implementation of fallback logic that mimics structural abstention. They differ entirely in scope, dealing with orthographic, structural, and localized data rather than abstract conceptual meaning29. The lesson for Embedded Semantics is that a global registry requires strict governance, immutability of assigned IDs, and clearly defined syntax rules akin to BCP 47 language tags28. The critical mistake to avoid is introducing breaking changes; CLDR's strict rule that deprecated attributes must remain for backward compatibility is absolutely essential for the longevity of a permanent conceptual registry32.

18. Feature Comparison Matrix

19. Similarities

Embedded Semantics is not developed in an intellectual vacuum; it shares its fundamental architectural DNA with several highly successful historical paradigms. With symbolic knowledge bases such as Wikidata, UMLS, and the Semantic Web, Embedded Semantics shares the absolute necessity of utilizing an abstract, registry-backed identifier (akin to a Q-node, CUI, or URI) that serves as an immutable anchor for conceptual meaning, entirely independent of the specific lexical strings used to describe it9. This shared commitment ensures that identity remains stable despite the evolutionary drift of language. With cross-lingual mapping frameworks like BabelNet and the Universal Networking Language (UNL), Embedded Semantics shares a deep commitment to establishing a multilingual interlingua, where core concepts successfully transcend the morphological constraints and syntactic barriers of individual human languages5. Furthermore, with contemporary vector databases and embedding architectures, Embedded Semantics shares the recognition that continuous latent spaces, dense vectors, and cosine similarity calculations represent the most computationally efficient and effective means currently available to perform broad-scale, zero-shot semantic retrieval over unstructured data.

20. Differences

Where Embedded Semantics diverges radically from the surrounding ecosystem is in its architectural synthesis and strict epistemological boundaries regarding the nature of identity. Unlike vector databases and large language models, Embedded Semantics flatly rejects the dense vector as the unit of identity. Vectors are inherently volatile, inextricably model-dependent, and prone to conflating evidence. In the Embedded Semantics framework, vectors are strictly classified as subordinate evidence pointing toward a stable Concept ID. Unlike traditional symbolic knowledge graphs such as Wikidata and the Semantic Web, Embedded Semantics natively incorporates dense vectors and machine learning representations as the primary mechanisms for retrieval and real-time evidence generation, rather than relying exclusively on explicit, human-curated logical triples and strict formal proofs19. Crucially, unlike all current systems analyzed, Embedded Semantics explicitly and formally models semantic residue1. It captures the stylistic, emotional, or highly contextual nuance that dense vectors capture but symbolic IDs inherently drop. Furthermore, it provides a formal, system-level framework for abstention, explicitly allowing an artificial intelligence to cryptographically state that it lacks sufficient, non-conflicting evidence to resolve an ambiguity, rather than forcing a probabilistic hallucination.

21. Ideas Already Well Solved Elsewhere

The Embedded Semantics framework is designed to integrate with, rather than reinvent, foundational technologies that have already achieved maturity. Graph traversal and formal logical reasoning are already handled perfectly by standards such as SPARQL, OWL, and commercial graph databases like Neo4j; Embedded Semantics relies on these for executing compositional relationships. The mathematical execution of Approximate Nearest Neighbor (ANN) search has been effectively commoditized by technologies like HNSW and enterprise vector databases (Pinecone, Milvus), which will serve as the physical evidence storage layer. The exhaustive mapping of named entities (such as specific cities, historical figures, and corporations) is already masterfully maintained by Wikidata and DBpedia, providing a baseline entity registry that does not need duplication. For specific, highly regulated domains, clinical semantics and taxonomies are already governed by UMLS and SNOMED CT8, while Unicode and CLDR definitively own the standardization of localized orthographic formatting29. Embedded Semantics maps to these resources rather than replacing them.

22. Areas Requiring Original Work

The realization of Embedded Semantics requires deep original engineering and theoretical advancement in several key areas. The most critical requirement is the development of the Registry and Mapping Protocol—a globally accessible, ultra-low-latency registry architecture capable of probabilistically mapping fluctuating vectors from arbitrary, distinct models (e.g., OpenAI, Anthropic, open-source variants) to stable Concept IDs. A second massive area of original work involves handling semantic residue. This requires developing a novel mathematical and symbolic framework capable of isolating and encoding the specific contextual "residue" (such as domain-specific tone or rhetorical sarcasm) that is left over when a raw text input is projected into a generalized Concept ID1. Furthermore, significant research is required in negative evidence modeling within latent space, creating training methodologies that allow systems to utilize vectors to explicitly query what a concept is not, effectively bounding the semantic space to prevent hallucinations. Finally, formalizing abstention thresholds requires the development of algorithmic confidence limits that dictate precisely when a system must declare explicit ambiguity rather than forcing a nearest-neighbor match.

23. Potential Research Contribution

Embedded Semantics offers a paradigm-shifting contribution to the fields of AI Alignment, Knowledge Representation, and Enterprise Architecture. By definitively decoupling identity from latent vectors, it provides the first viable pathway to solve the escalating "LLM upgrade problem," a scenario where businesses currently lose all of their semantic indexing and operational continuity every time a foundational model is updated. Furthermore, by forcing opaque AI systems to map their latent representations to verifiable, discrete Concept IDs, Embedded Semantics introduces a cryptographically auditable, version-controlled provenance trail directly into deep learning workflows. This represents a crucial, currently missing requirement for achieving regulatory compliance—such as adhering to the EU AI Act—and establishing truly trustworthy AI systems35.

24. Areas Where Claims Must Remain Modest

While the architectural implications are vast, claims surrounding Embedded Semantics must remain strictly bounded by empirical reality. First, Embedded Semantics is an infrastructure standard; it is not a cognitive architecture, and it does not inherently create General Artificial Intelligence (AGI) or bestow machines with actual consciousness or "understanding." It merely organizes existing intelligence effectively. Second, claims regarding latency must remain modest. Decoupling identity from vectors inherently introduces a lookup and resolution step into the retrieval pipeline, meaning raw computational speed cannot perfectly match that of a bare vector database performing pure cosine similarity without secondary resolution. Finally, claims regarding the complete automation of ontology generation must be tempered. While LLMs can efficiently suggest mappings and surface relationships, generating the definitive boundaries and ontological rigor of a new Concept ID still fundamentally requires human-in-the-loop curation or domain expert consensus to prevent drift.

25. Recommended Terminology

26. Recommended Project Narrative

1\. For General Technical Readers: Today, AI systems define the meaning of the world by converting words into massive lists of numbers called vectors. But every time a new AI model is built or updated, those numbers change entirely, meaning the AI essentially forgets or arbitrarily shifts its understanding of the world. Embedded Semantics fixes this fundamental flaw by giving every single concept a permanent, stable ID. The vectors are no longer the meaning; they are just temporary clues used to find that stable ID. This means AI can finally possess a stable, versioned memory that doesn't break every time the underlying technology updates. 2\. For Software Engineers: Current vector databases treat the mathematical embedding as the primary key of the system. This is a massive architectural flaw because vectors are highly volatile and inextricably coupled to specific model versions, causing severe vendor lock-in and destroying backward compatibility upon upgrade. Embedded Semantics introduces a vital abstraction layer: vectors are demoted to serve merely as search indexes (evidence), while the primary key becomes a stable, registry-backed Concept ID. This definitively decouples your application's semantic and business logic from the underlying embedding model. 3\. For ML Researchers: Current latent representations suffer severely from semantic conflation, anisotropic spaces, and entangled features, making explicit negative evidence and systemic abstention mathematically difficult to enforce. By mapping latent model outputs to a stable, overarching symbolic registry, Embedded Semantics bounds the representation space. It captures and isolates semantic residue via sparse priors or explicit relational tagging2, allowing for true multi-modal, model-agnostic semantic alignment without catastrophic forgetting during iterative fine-tuning. 4\. For Linguists: While classical systems like WordNet rely on rigid lexical relations, and FrameNet relies on exhaustive manual role annotation3, modern LLMs rely purely on statistical co-occurrence, ignoring linguistic structure entirely. Embedded Semantics unites these disparate approaches. It utilizes statistical vectors to handle the fuzzy, distributional nature of human language, but deliberately anchors them to discrete, language-agnostic Concept IDs. It formally models polysemy through 'explicit ambiguity' and rigorously captures the pragmatic, stylistic nuances often lost in translation as 'semantic residue'10. 5\. For Knowledge-Graph Researchers: The Semantic Web community has the perfect structural foundation—URIs, RDF, and strict provenance—but completely lacks the ability to ingest unstructured data dynamically via zero-shot probabilistic retrieval. Embedded Semantics bridges this historic gap. It treats your URIs as the ultimate ground truth (Concept IDs) and utilizes dense vectors as the probabilistic evidence layer required to map unstructured, messy text into your highly structured graph. It is the missing neuro-symbolic interface. 6\. For Standards Professionals: Global standards absolutely require immutability, rigorous versioning, and traceable provenance. Current AI models operate as opaque black boxes that violate all three of these foundational requirements. Embedded Semantics provides an infrastructure akin to Unicode or CLDR29, but engineered for semantic meaning rather than just character encoding. By forcing AI models to resolve their outputs to stable, version-controlled Concept IDs, we reintroduce auditability, safety bounds, and seamless interoperability into machine learning systems.

27. Homepage Content Architecture

28. About Page

The About Page must focus heavily on the philosophical and epistemological foundation of the project. It should articulate the fundamental error of modern AI in treating statistical representation as ground-truth identity. The narrative must frame Embedded Semantics not as a new product to buy, but as a necessary infrastructural evolution. It should explicitly acknowledge its intellectual heritage, stating that it stands on the shoulders of the Semantic Web's rigid identities19, the cross-lingual mapping of translation interlinguas5, and the generalized power of deep learning, while purposefully correcting the respective architectural blind spots of each.

29. Research Page

The Research Page will serve as the academic hub of the project, highlighting the formal definitions and ongoing studies surrounding semantic residue, explicit abstention, and negative evidence. It must link to ongoing empirical experiments demonstrating how mapping latent spaces to Concept IDs measurably reduces hallucination rates in LLMs and enables zero-shot cross-model interoperability. The page should prominently feature literature on sparse autoencoders separating stylistic signals from semantic content1, demonstrating that Embedded Semantics is grounded in cutting-edge representation learning research.

30. Architecture Page

The Architecture Page will provide the definitive technical specifications of the framework, broken down into four core modules:

1. The Evidence Layer: How vectors from disparate models are ingested, normalized, and treated probabilistically.

2. The Mapping Protocol: The specific algorithmic thresholds and confidence scoring mechanisms used to resolve vectors to Concept IDs.

3. The Global Registry: The exact schema for a Concept ID, including its ID structure, Multilingual Labels, Version History tracking, and Provenance Metadata constraints.

4. The Semantic Graph: How compositional relationships between Concept IDs are structured and queried using standard graph protocols.

31. FAQ

32. Comparison Pages

The site should feature dedicated comparison pages to ensure accurate positioning against adjacent technologies.

• Embedded Semantics vs. Vector Databases: Focus on the distinction between true Semantic Identity versus probabilistic indexing.
• Embedded Semantics vs. Semantic Web (RDF/OWL): Focus on the necessity of probabilistic vector evidence in handling messy, unstructured data versus the fragility of strict logical triples.
• Embedded Semantics vs. WordNet / UMLS: Focus on the transition from static, manually curated lexical structures to continuous, multilingual, dynamically updated latent mapping, while retaining their mastery of stable IDs.

33. Proposed White Paper Structure

34. Proposed Technical Documentation Structure

35. Diagram Library

36. 25 Strongest Content Pieces to Publish

37. Research Roadmap

The development of Embedded Semantics requires a phased, rigorous research roadmap to move from architectural theory to enterprise deployment.

• Phase 1 (Foundation): Formalize the strict Concept ID schema and build the initial, highly performant registry APIs required to handle millisecond resolution queries.
• Phase 2 (Mapping Middleware): Develop open-source middleware designed to sit between standard LLM outputs (e.g., OpenAI, HuggingFace) and the Concept Registry. This phase focuses heavily on optimizing the configurable confidence thresholds required to translate vectors into IDs reliably.
• Phase 3 (Residue & Ambiguity Frameworks): Implement the formal mathematical and symbolic logic for explicitly storing semantic residue and triggering abstention states during Retrieval-Augmented Generation (RAG) pipelines.
• Phase 4 (Ecosystem Integration): Integrate natively with major Vector Databases (Pinecone, Milvus) and Graph Databases (Neo4j), offering a seamless, drop-in replacement for standard embedding-based primary keys to drive enterprise adoption.

38. Open Questions

• Initial Seeding: How is the initial seed of Concept IDs generated to achieve critical mass? Should the system bootstrap from highly curated resources like Wikidata and UMLS, or start from a foundational blank slate driven entirely by latent clustering?
• Threshold Mathematics: What is the exact mathematical threshold or algorithm used to determine when a vector constitutes definitive "positive evidence" versus when the system must trigger an "abstain" state due to ambiguity?
• Residue Standardization: How can semantic residue be standardized and encoded efficiently so it does not degrade into a bloated dumping ground for unstructured data?
• Global Governance: Who governs the global registry of Concept IDs to prevent malicious altering of versioned meaning, and should this be a centralized consortium or a decentralized protocol?

39. Annotated Bibliography

The intellectual heritage of Embedded Semantics draws upon several distinct, historically isolated disciplines, synthesizing their insights while addressing their systemic failures. In the realm of the Semantic Web and Ontologies, Berners-Lee’s vision of machine-readable data established the absolute necessity of URIs as stable identifiers, utilizing RDF and OWL to embed semantics directly into data structures19. While highly precise and structurally sound, these symbolic frameworks ultimately struggled with the fluidity and scale of natural language processing. Within the biomedical domain, the UMLS (Unified Medical Language System) and SNOMED CT provided the archetypal demonstration of successfully decoupling conceptual identity from string representation. The UMLS Semantic Network established Concept Unique Identifiers (CUIs) and a rigorous network of Semantic Types and Relations8. This proven, operational capability to maintain versioned, highly stable IDs across disparate vocabularies directly informs the Embedded Semantics registry model. Efforts to bridge global linguistic divides, such as the Universal Networking Language (UNL), demonstrated the viability of a semantic interlingua. By representing knowledge as complex hypergraphs of Universal Words rather than relying on syntax, UNL decoupled core consensual meaning from language-specific morphology5. Similarly, BabelNet successfully merged lexicographic data (WordNet) and encyclopedic knowledge (Wikipedia) into a massive multilingual semantic network, highlighting the absolute importance of provenance tracking and confidence scoring in cross-lingual mappings6. Within cognitive linguistics, FrameNet advanced the concept of Frame Semantics, proving empirically that meaning is deeply compositional and contextual. By formally separating core frame elements from non-core, peripheral elements3, FrameNet provided a theoretical foundation for what Embedded Semantics formalizes computationally as semantic residue—the peripheral, contextual nuance that accompanies a core semantic assertion. Finally, recent advancements in the analysis of Large Language Models and Sparse Autoencoders (SAEs) have highlighted the severe limitations of dense continuous vectors. Researchers utilizing frameworks like GLASS have demonstrated that dense representations suffer from semantic contamination and "residue," where stylistic signals become dangerously entangled with core meaning1. Interventions designed to explicitly scale text embedding variance and untangle semantic residue2 validate the core premise of Embedded Semantics: vectors are volatile, highly entangled evidence that require an overarching, stable symbolic architecture to ensure precision, safety, and reliability.

Claims Embedded Semantics Can Defend Today

1. Dense vectors are inherently volatile and inextricably tied to specific model architectures, making them categorically unsuitable as primary keys for long-term semantic knowledge storage.

2. Existing commercial vector databases lack native architectural mechanisms for versioning meaning, tracking evidentiary provenance, and handling explicit systemic abstention.

3. Symbolic knowledge graphs (like Wikidata and UMLS) possess the correct identity architecture through stable IDs, but they entirely lack the zero-shot generalization capabilities provided by dense neural retrievers.

4. By structurally decoupling the Concept ID from the Evidence Vector, enterprise organizations can change their underlying embedding models without destroying their established semantic knowledge bases.

5. Explicitly modeling ambiguity and abstention prevents the forced-choice hallucinations that are mathematically guaranteed in standard nearest-neighbor vector search.

Claims That Require Experimental Evidence

1. Real-time mapping of latent vectors to Concept IDs can be executed without introducing latency overheads that break high-speed, enterprise-scale RAG applications.

2. Semantic residue can be mathematically or symbolically encoded efficiently without causing unacceptable bloat to the storage and retrieval architecture.

3. Negative evidence modeling via vector constraints can statistically eliminate LLM hallucinations in uncontrolled production environments.

4. A unified registry of Concept IDs can adequately capture the extreme nuance of all human languages without falling into the rigid, brittle taxonomic traps that caused WordNet to fail at scale.

Claims Embedded Semantics Should Not Make

1. Embedded Semantics is a new type of LLM, neural network, or embedding model that outperforms OpenAI or Anthropic in natural language generation tasks.

2. The framework completely eliminates the need for human-in-the-loop curation of knowledge boundaries and ontological relationships.

3. The architecture achieves Artificial General Intelligence (AGI) or somehow imbues machines with true cognitive "understanding" or consciousness.

4. Embedded Semantics replaces vector databases entirely (it strictly enhances them and acts as the vital semantic governance layer directly above them).

Works cited

1. From Profiles to Steering Vectors: Global Sparse Priors and Local Semantic Calibration for Personalized Text Generation \- arXiv, https://arxiv.org/html/2607.21620v1

2. From Profiles to Steering Vectors: Global Sparse Priors and Local Semantic Calibration for Personalized Text Generation | alphaXiv, https://www.alphaxiv.org/abs/2607.21620

3. FrameNet \- Wikipedia, https://en.wikipedia.org/wiki/FrameNet

4. FrameNet Lexical Databases: Knowledge Graph Extraction — Case Study \- Medium, https://medium.com/@jolalf/framenet-lexical-databases-knowledge-graph-extraction-case-study-9def7c3fab1a

5. English to UNL (Interlingua) Enconversion \- CSE IITB, https://www.cse.iitb.ac.in/\~damani/papers/LTC09/unlLTC09.pdf

6. Converting BabelNet as Linguistic Linked Data \- Best Practices for Multilingual Linked Open Data Community Group, https://www.w3.org/community/bpmlod/wiki/Converting\_BabelNet\_as\_Linguistic\_Linked\_Data

7. Guidelines for Linguistic Linked Data Generation: Multilingual Dictionaries (BabelNet) \- W3C, http://www.w3.org/2015/09/bpmlod-reports/multilingual-dictionaries/

8. Semantic Network \- UMLS® Reference Manual \- NCBI Bookshelf, https://www.ncbi.nlm.nih.gov/books/NBK9679/

9. Relationship Structures and Semantic Type Assignments of the UMLS Enriched Semantic Network | Journal of the American Medical Informatics Association | Oxford Academic, https://academic.oup.com/jamia/article/12/6/657/693225

10. From Profiles to Steering Vectors: Global Sparse Priors and Local Semantic Calibration for Personalized Text Generation \- arXiv, https://arxiv.org/pdf/2607.21620

11. CMC | Free Full-Text | Do LLMs Know When Evidence is Insufficient? An Evidence Sufficiency Benchmark for Answer-Abstention Calibration in Retrieval-Augmented Generation, https://www.techscience.com/cmc/v89n1/68467/html

12. Word Sense Disambiguation with Wikipedia Entities: A Survey of Entity Linking Approaches \- PMC, https://pmc.ncbi.nlm.nih.gov/articles/PMC12939010/

13. Managing Provenance and Versioning for an (Evolving) Dictionary in Linked Data Format, https://www.researchgate.net/publication/330201482\_Managing\_Provenance\_and\_Versioning\_for\_an\_Evolving\_Dictionary\_in\_Linked\_Data\_Format

14. Toward data lakes as central building blocks for data management and analysis \- Frontiers, https://www.frontiersin.org/journals/big-data/articles/10.3389/fdata.2022.945720/full

15. FrameNet at 25 | International Journal of Lexicography \- Oxford Academic, https://academic.oup.com/ijl/article/37/3/263/7708430

16. FrameNet: Theory and Practice, https://ids-pub.bsz-bw.de/frontdoor/deliver/index/docId/5416/file/Johnson\_Petruck\_Baker\_Ellsworth\_Ruppenhofer\_Fillmore\_FrameNet\_Theory\_and\_Practice\_2003.pdf

17. FrameNet Resource Grammar Library for GF \- arXiv, https://arxiv.org/pdf/1406.6844

18. Is FrameNet usage-based? Predicting frame element coreness with information theory and gradient boosting | Language and Cognition, https://www.cambridge.org/core/journals/language-and-cognition/article/is-framenet-usagebased-predicting-frame-element-coreness-with-information-theory-and-gradient-boosting/FA823BBD1BB57B31D685B7BE549682CB

19. Semantic Web \- Wikipedia, https://en.wikipedia.org/wiki/Semantic\_Web

20. THE ROLE OF EMBEDDED SEMANTICS 1 Introduction Pervasive computing \[1, 2\] \- River Publishers, https://journals.riverpublishers.com/index.php/JMM/article/download/4779/3501/13695

21. Effectiveness of UMLS semantic network as a seed ontology for building a medical domain ontology | Aslib Journal of Information Management \- Emerald Insight, https://www.emerald.com/ajim/article/60/1/32/37907/Effectiveness-of-UMLS-semantic-network-as-a-seed

22. UMLS Semantic Network \- National Library of Medicine, https://www.nlm.nih.gov/research/umls/knowledge\_sources/semantic\_network/index.html

23. IHTSDO/snomed-expression-constraint-language: Formal syntax and valid examples for each version of ECL \- GitHub, https://github.com/IHTSDO/snomed-expression-constraint-language

24. ECLed– a tool supporting the effective use of the SNOMED CT Expression Constraint Language \- PMC, https://pmc.ncbi.nlm.nih.gov/articles/PMC12777381/

25. Universal Networking Language Overview | PDF \- Scribd, https://www.scribd.com/document/657494123/NLP-Unit-5

26. janardhan: Semantic Textual Similarity using Universal Networking Language graph matching \- ACL Anthology, https://aclanthology.org/S12-1098.pdf

27. CLDR 48 Release Note \- Unicode CLDR Project, https://cldr.unicode.org/downloads/cldr-48

28. Language Tags and Locale Identifiers for the World Wide Web \- W3C, https://www.w3.org/TR/ltli/

29. CLDR 46 Release Note \- Unicode CLDR Project, https://cldr.unicode.org/downloads/cldr-46

30. JSON endpoints for CLDR exemplar data by locale tag \- GitHub, https://github.com/googlefonts/exemplar

31. Unicode Locale Data Markup Language (LDML), https://www.unicode.org/reports/tr35/

32. Updating DTDs \- Unicode CLDR Project, https://cldr.unicode.org/development/updating-dtds

33. Unicode Locale Data Markup Language (LDML) Part 2: General, https://www.unicode.org/reports/tr35/tr35-70/tr35-general.html

34. IETF language tag \- Wikipedia, https://en.wikipedia.org/wiki/IETF\_language\_tag

35. The Semantic Layer: Architecture, Components, and the Foundation for Trustworthy AI, https://software.strategy.com/blog/the-semantic-layer-architecture-components-and-the-foundation-for-trustworthy-ai

36. Picking reference events from tense trees: a formal, implementable theory of English tense-aspect semantics \- ResearchGate, https://www.researchgate.net/publication/234793385\_Picking\_reference\_events\_from\_tense\_trees\_a\_formal\_implementable\_theory\_of\_English\_tense-aspect\_semantics

37. The Workshop Programme \- ACL Anthology, https://aclanthology.org/www.mt-archive.info/00/LREC-2002-WS-UNL.pdf