Multilingual Semantic Representation and Cross-Lingual Alignment: A Registry-First Architecture

Full report

On this page

1. Executive Summary

The pursuit of a universal semantic representation has historically oscillated between two theoretical extremes. The first assumes that all human languages can be losslessly mapped onto a single, isomorphic ontological framework, a premise that underpins much of contemporary commercial machine translation. The second, rooted in linguistic relativity, argues that languages are fundamentally incommensurable, rendering perfect translation impossible. The research presented in this report navigates a synthesis of these paradigms for EmbeddedSemantics.com. It operates on the foundational philosophy that while diverse linguistic expressions can converge toward stable, registry-backed concepts, cross-lingual translation is inherently lossy. This report provides an exhaustive investigation into multilingual semantic representation. It rigorously analyzes the geometric properties of cross-lingual vector spaces, the complexities of semantic universals, and the structural realities of untranslatability. The research strictly rejects the paradigm of utilizing "translation-to-English" as a permanent semantic representation. English-centric models enforce an epistemological bottleneck, actively erasing critical grammatical, cultural, and pragmatic distinctions—such as evidentiality, socio-linguistic honorifics, shape-based classifiers, and aspectual nuances—that are either non-existent or mapped entirely differently in the English language. To construct a shared semantic layer without destroying language-specific granularity, this report proposes a highly structured architecture for "Semantic Residue." Semantic residue captures the meaning left behind when an expression is mapped to a shared primary Concept. By evaluating contemporary multilingual embedding models (e.g., LaBSE, LASER, BGE-M3), geometric pathologies (specifically anisotropy and centrality-driven hubness), and classical interlingua systems, the research demonstrates that vector embeddings serve merely as candidate evidence for semantic proximity. They do not, and geometrically cannot, define semantic identity. The report outlines an advanced taxonomy for semantic residue, provides mathematical guidelines for constructing language-independent concept prototypes (favoring medoids and Cross-Domain Similarity Local Scaling over naive centroids), and establishes an evaluation methodology to benchmark semantic alignment while preserving the rich, compositional semantics of the world's languages.

2. What Cross-Language Semantic Equivalence Actually Means

In computational linguistics, cross-language semantic equivalence is frequently conflated with distributional similarity or translation probability. However, true semantic equivalence is rarely a mathematical isomorphism. It requires alignment across three distinct vectors of meaning: extensionality (the real-world referent or truth conditions), intensionality (the conceptual criteria, sense, or cognitive framing), and pragmatic force (the social, performative, and contextual function of the utterance). When an expression in one language is aligned with an expression in another, they may share extensional equivalence but diverge significantly in their intensional or pragmatic dimensions. Semantic equivalence is therefore not a binary state but a spectrum of convergence. A formal greeting and an informal greeting may both map to the same extensional event—the act of welcoming an interlocutor—but their pragmatic forces differ entirely. If a system enforces strict equivalence, it deletes the social hierarchy encoded in the formal greeting. Embedded Semantics conceptualizes equivalence not as absolute interchangeability, but as convergence upon a shared "Concept Identity." The shared concept acts as a stable ontological anchor within a registry. The divergences—the nuances that make the expressions distinct in their native languages—are preserved rather than discarded. This framework acknowledges that forcing complete equivalence across languages invariably requires the deletion of culturally and grammatically specific information. Furthermore, semantic universals (concepts that exist identically in every human language) are exceedingly rare. Most concepts are subject to varied lexicalization patterns, meaning that a robust semantic representation must tolerate asymmetry, partial overlaps, and lexical gaps without forcing a false equivalence.

3. Translation versus Semantic Identity

Translation is an act of communicative approximation. A translator's objective is to produce a target-language expression that fulfills the communicative intent of the source-language expression, relying on compensatory strategies when direct semantic equivalence is impossible. Semantic identity, conversely, is the theoretical state wherein two expressions share exact ontological, conceptual, and pragmatic footprints. Modern machine translation and cross-lingual representation learning blur this crucial distinction. When a neural machine translation model translates the Javanese krama inggil (high formal) and ngoko (informal) second-person pronouns both as the English "you," it succeeds in translation but fails in semantic identity1. The hierarchical social dynamic—the very core of the Javanese expression's pragmatic utility—is eradicated. Similarly, when a system translates a term carrying deep cultural weight, the translation serves as a functional bridge but leaves behind a vast amount of retained, lost, and added meaning. Therefore, translation-to-English cannot serve as the permanent semantic representation for a global architecture. English lacks the morphological and syntactic machinery to encode evidentiality (how information was acquired), complex aspectual distinctions, or intricate sociolinguistic registers. Relying on English as an interlingua forces all global concepts through an Anglo-centric filter, stripping expressions of their semantic residue and reducing them to the lowest common denominator of meaning. Semantic identity must be maintained independently of any single language's lexical constraints.

4. Multilingual Embedding Research

The foundational premise of multilingual embedding spaces is that semantically similar sentences, regardless of language, should be located near one another in a shared, high-dimensional vector space. Several prominent architectures have attempted to achieve this, each with distinct optimization strategies and inherent limitations regarding semantic identity. Current cross-lingual representation learning relies heavily on models like LASER (Language-Agnostic SEntence Representations), which utilizes a BiLSTM encoder-decoder architecture trained on massive parallel corpora3. LASER is optimized for translation, establishing semantic symmetry by forcing source and target sentences into the same vector space. However, it struggles with fine-grained semantic distinctions within a single language. Conversely, LaBSE (Language-agnostic BERT Sentence Embedding) employs contrastive learning on translation pairs, pulling translations closer and pushing non-translations apart3. While LaBSE excels at cross-lingual sentence alignment and zero-shot transfer, its contrastive objective deliberately erases semantic nuances (such as register, aspect, or modality) that differentiate a source sentence from its translation in order to maximize vector cosine similarity4. Multilingual knowledge distillation, popularized by Reimers and Gurevych, trains a "student" model to mimic a "teacher" model's outputs across translated sentences6. While this efficiently aligns vector spaces, it propagates the English-centric biases of the original teacher model, ensuring that the multilingual space inherits the semantic blind spots of English8. The current state-of-the-art includes models like Multilingual E5 and BGE-M34. BGE-M3 is optimized for multi-granularity, multi-functional dense retrieval, achieving remarkable zero-shot performance. Yet, these models still suffer from input constraints and rely heavily on the availability of massive multilingual training data11. While embeddings provide excellent candidate evidence for semantic alignment—allowing systems to retrieve candidate matches across languages—they do not define semantic identity. The contrastive losses used to train these models treat all translation pairs as perfectly synonymous, systematically destroying the semantic residue that Embedded Semantics seeks to preserve.

5. Interlingua Research

The concept of an interlingua—a language-neutral representation mediating between natural languages—has deep historical roots in machine translation. The most prominent computational effort in this domain is the Universal Networking Language (UNL), initiated by the United Nations University12. UNL represents sentences as hypergraphs, where nodes are "Universal Words" (UWs) and edges are semantic relations (e.g., agent, instrument, time, place)14. Attributes (e.g., @past, @interrogative, @def) modify the UWs to capture grammatical and modal information14. UNL successfully decouples semantic relations from surface syntax, providing a robust ontological framework for language-neutral representation. However, UNL's reliance on English headwords for its UWs introduces subtle lexical biases, anchoring the interlingua to the semantic boundaries of English vocabulary. Furthermore, rigid symbolic interlinguas often struggle with concepts that resist decomposition into predefined primitives, such as idioms or highly specific cultural practices. While Embedded Semantics shares UNL's graph-based, relation-heavy philosophy, it diverges by utilizing stable, registry-backed Concept identities rather than relying on English-derived vocabulary, and by formally integrating continuous vector spaces alongside discrete graph edges to manage ambiguity and compositional semantics.

6. Lexical Gaps

A lexical gap occurs when a language lacks a single lexical item to express a concept that is lexicalized in another language. These gaps demonstrate the fallacy of one-to-one semantic interchangeability and highlight the reality of untranslatable concepts. The Sapir-Whorf hypothesis, or linguistic relativity, provides a framework for understanding these gaps. While linguistic determinism (the strong form positing that language strictly limits thought) is largely rejected by modern cognitive scientists, linguistic relativity (the weak form) is widely accepted: language influences cognition and facilitates the conceptualization of specific cultural and environmental paradigms17. Lexical gaps often represent concepts that are highly salient in one culture but peripheral in another. For instance, the German Schadenfreude (pleasure derived from another's misfortune) or the Japanese Tsundoku (buying books and leaving them unread) represent complex, compositional events bundled into single lexical items20. When translated into a language with a lexical gap, these must be expanded into phrasal descriptions. This expansion changes the compositional semantics and the cognitive load of the expression. In a shared semantic layer, the primary concept must accommodate the original unitary concept, while the architecture must acknowledge that other languages will rely on compositional semantic structures to approximate it, resulting in added or lost meaning during alignment.

7. Cultural and Pragmatic Meaning

Meaning extends far beyond the literal decoding of words into the realm of pragmatics—how context, culture, and speaker intent shape interpretation. Idioms, metaphors, conversational implicatures, and culturally specific concepts frequently resist cross-lingual alignment because their extensional meaning is inextricably linked to localized social contracts. For example, the Indonesian term Jayus refers to a joke told so poorly that one cannot help but laugh22. The pragmatic force of this word relies on specific social expectations of humor, failure, and communal reaction. Similarly, politeness strategies differ vastly. Direct translations of requests from English into Russian or Hebrew often sound overly polite, distant, or evasive to native speakers, while translations in the reverse direction may sound abrupt, demanding, or rude. Semantic representation must account for these pragmatic dimensions. When expressions converge on a concept prototype, the cultural specificity and pragmatic force must not be stripped away as statistical noise. Instead, this information must be encoded as highly structured semantic residue, acknowledging that the mapping of a specific locale's idiom to a universal concept is an approximation that leaves behind critical communicative intent.

8. Grammar-Carried Meaning

Perhaps the most significant loss in translation-to-English models is grammar-carried meaning. Grammatical categories force speakers to attend to specific aspects of reality, embedding these observations into the morphological fabric of the utterance. When translated into languages lacking these categories, this mandatory cognitive framing evaporates. Evidentiality: In languages like Tuyuca and Turkish, speakers must grammatically encode their source of information24. Tuyuca requires verbs to indicate whether an event was seen directly, heard, inferred, or learned via hearsay, maintaining up to five levels of epistemological stance27. An English translation simply stating "He played soccer" completely erases the Tuyuca speaker's evidentiary basis. Aspect and Tense: Russian strictly enforces aspectual distinctions (perfective versus imperfective), dictating whether an action is viewed as a completed whole or as an ongoing/habitual process29. An English simple past tense often fails to capture this distinction entirely. Tense systems also vary widely; some languages utilize spatial deictic past tenses that intertwine temporal distance with physical space. Classifier Systems: Tzeltal, a Mayan language, uses highly specific sortal numeral classifiers that categorize objects by physical properties (e.g., shape, orientation, rigidity) when counting30. Counting "two objects" requires specifying whether the objects are flat, round, or elongated. This enforces a continuous cognitive categorization of the physical world that disappears in languages without numeral classifiers. Honorifics and Kinship: Javanese and Korean speech levels require speakers to encode their social status relative to the listener and the referent in almost every word of the sentence, establishing a strict social ontology1. Furthermore, kinship systems in languages like Pitjantjatjara or Chinese divide family relations into highly specific categories (e.g., maternal vs. paternal uncles, relative age of siblings) that broad English terms like "uncle" or "brother" completely obfuscate33. A robust multilingual architecture must treat these grammatical categories not as syntactic noise, but as core semantic features that define the exact coordinates of the expression in the global semantic space.

9. Semantic Residue

Semantic residue is defined as the meaningful information left over when an expression is mapped to its closest language-independent primary concept34. Because translation is not a perfect bijection, forcing a native expression into a primary concept prototype inevitably leaves a remainder. This remainder encapsulates the retained meaning, the potentially lost meaning, and the potentially added meaning inherent in cross-lingual alignment. Consider the French expression dépaysement. Its primary concept mapping might be HOMESICKNESS or DISORIENTATION. However, its true semantic residue includes the specific trigger (being in a foreign country) and a subtle blend of exhilaration and alienation22. If this residue is not formally captured, the distinctiveness of the French expression is lost, and the semantic architecture regresses to a naive translation dictionary. Model theoretic semantics often struggles with this "conceptual leakage," failing to account for the rich, polysemous conceptual content that escapes strict logical reference35. Embedded Semantics must therefore operationalize residue as a first-class citizen in its data architecture, explicitly modeling what is left behind during conceptual alignment.

10. Proposed Residue Taxonomy

To rigorously model semantic residue without violating the registry-first philosophy, residue must not be modeled arbitrarily. Evaluating the tradeoffs among secondary concepts, attributes, relations, frames, features, graph edges, and compositional semantic structures yields the following recommended taxonomy for Embedded Semantics:

Recommendation: Implement a hybrid model relying primarily on Attributes for grammar-carried and sociolinguistic meaning, and Relational Edges to secondary concepts for added or lost conceptual meaning. This aligns perfectly with a registry-first philosophy, ensuring that semantic residue is systematically categorized rather than left as unstructured text.

11. Concept Prototypes

In cognitive linguistics, prototype theory posits that semantic categories are organized around a central, prototypical exemplar, with other members radiating outward based on family resemblance37. In the context of Embedded Semantics, a "Concept Prototype" serves as the stable registry identity representing the intersection of aligned cross-lingual expressions. Constructing this prototype theoretically involves identifying the central tendency of a concept across multiple languages. The prototype is not tied to any single language; rather, it is an abstract anchor point. If English "dog," Spanish "perro," and Japanese "inu" converge on the concept of DOG, the prototype is the geometric and semantic center of these aligned representations. However, mathematically deriving this prototype from vector embeddings introduces significant geometric challenges that, if ignored, will fundamentally corrupt the semantic registry.

12. Multilingual Centroids

When constructing a concept prototype from reviewed expressions in multiple languages, simple vector averaging (computing a normalized mean vector or weighted centroid) is often intuitively appealing but geometrically perilous. Anisotropy and Hubness: High-dimensional spaces generated by Transformer models exhibit extreme anisotropy: vectors collapse into a narrow, low-dimensional cone, failing to occupy the ambient space uniformly39. Consequently, average pairwise cosine similarities are unnaturally high. When vectors are averaged, the resulting centroid is pushed toward the densest part of this cone. This triggers "centrality-driven hubness"40. A hub is a point that appears as the nearest neighbor to a disproportionately large number of unrelated vectors. Recent studies prove that hubness—not anisotropy directly, centroid drift, or magnitude—is the dominant causal driver of cross-lingual retrieval asymmetry, accounting for 49.5% of explained variance in retrieval failures40. Averaging vectors blindly creates a meaningless "super-hub" that matches with everything and nothing, destroying embedding-space alignment. Valid Approaches to Prototypes: To maintain scientific validity when generating prototypes, the following approaches must be evaluated:

• Cross-Domain Similarity Local Scaling (CSLS): CSLS is absolutely critical. It penalizes vectors in dense regions (hubs) and expands the space in sparse regions, correcting the asymmetry caused by hubness42. Any distance calculation to a prototype must use CSLS rather than raw cosine similarity.
• Medoids over Centroids: Instead of creating a synthetic averaged vector (a normalized mean), the system should select the medoid—the actual, attested vector of an expression that has the minimal average distance to all other aligned expressions in the concept cluster. This prevents the prototype from drifting into empty space or hub regions.
• Robust Centroids (Top-PC Removal): If a mathematical centroid is required, removing the top Principal Components (which dominate the anisotropic cone) before averaging can yield a more isotropic, representative position39.
• Prototype Spread and Covariance: The system must measure the covariance and prototype spread of the cluster. If the spread is too wide, averaging vectors is scientifically meaningless, indicating that the expressions do not actually share a stable semantic identity.

13. Multiple-Sense Problems

Polysemy degrades the validity of multilingual centroids and wreaks havoc on embedding-space alignment. If the English word "bank" is aligned with the French "banque" (financial institution) and the French "rive" (river edge), computing a singular centroid of all expressions associated with "bank" places the prototype in a semantic dead zone halfway between finance and nature37. Standard embeddings struggle heavily with sense disambiguation in highly anisotropic spaces39. An architecture must allow for multiple prototypes per concept through the use of Sense Clusters. Expressions must be clustered by their contextual representations prior to alignment. If a concept spans disjoint regions of the semantic space across languages, it must be explicitly split into distinct sub-concepts in the registry.

14. Language-Balancing Problems

Constructing a language-independent prototype from empirical expressions risks severe demographic skew. If a registry contains 10,000 English expressions, 5,000 Spanish expressions, and 50 Tuyuca expressions for a given concept, an unweighted centroid will simply result in an Anglo-European vector, implicitly forcing English as the semantic representation. To mitigate this, the architecture must employ Language-Family Weighting. Averages or medoid calculations must be weighted not by the volume of expressions, but by language family (e.g., Indo-European, Sino-Tibetan, Niger-Congo). Utilizing language-balanced prototypes prevents the resulting space from being biased toward the highest-resource language47. Furthermore, strict outlier detection must be implemented to ensure that a mistranslated expression in a low-resource language does not artificially drag the weighted prototype away from the true semantic center.

15. Failure Modes

Relying on embeddings as the sole arbiter of semantic identity introduces several critical failure modes:

1. Word Inversion: Embeddings often place antonyms (e.g., "hot" and "cold") in close proximity because they share identical syntactic contexts and distributional footprints37.

2. Topological Collapse: Dense embeddings struggle to capture the strict logical relations and compositional semantics between parts of a sentence, often collapsing complex relational data into a generic topical vector.

3. Cross-Lingual Hallucination: Contrastive learning models (like LaBSE) will force an English idiom and its literal, non-idiomatic translation in Arabic into the same vector space, even if the literal Arabic translation is entirely meaningless to a native speaker4. This hallucination creates false semantic alignments based on training objectives rather than semantic reality.

16. Evaluation Methodology

Embedded Semantics must not rely solely on standard Information Retrieval metrics (such as Recall@k or Mean Reciprocal Rank) to validate its architecture. Evaluation must explicitly measure semantic retention. Methodological steps for evaluation:

1. Residue Preservation Scoring: When mapping a source expression to a Concept and subsequently retrieving it in a target language, measure whether the original pragmatic and grammatical markers (e.g., formal register, past tense, evidentiality) can be successfully reconstructed or compensated for using the semantic residue.

2. Asymmetry Auditing: Track mutual nearest-neighbor rates between language pairs. If expression A retrieves expression B, but B retrieves expression C, the semantic alignment is geometrically unstable and indicates hubness interference40.

3. Human-in-the-loop Adjudication: Utilize native speaker evaluators to verify if the "lost meaning" and "added meaning" identified by the system accurately match human intuition regarding translation fidelity.

17. Recommended Benchmark Categories

To properly benchmark multilingual semantic representation within a registry-first philosophy, Embedded Semantics should adopt the following specific evaluation categories, transcending standard MTEB (Massive Text Embedding Benchmark) datasets:

• Grammatical Fidelity: Testing the precise preservation of evidentiality, complex aspect, and shape-based classifier usage.
• Pragmatic Alignment: Evaluating the mapping of politeness strategies, honorifics, and indirect speech acts across vastly different social hierarchies.
• Lexical Gap Resolution: Measuring how accurately a single culturally specific word maps to a compositional phrase without loss of core extensional meaning.
• Hubness Resistance: Quantifying the degree to which a derived concept prototype avoids becoming a universal nearest neighbor for unrelated queries.

18. Examples Across at Least 15 Languages

The following 30 examples demonstrate the absolute necessity of modeling semantic residue when mapping expressions to shared primary concepts. They illustrate situations where expressions are strong translations but are not semantically identical.

19. Implications for Embedded Semantics

The findings of this research strongly validate Embedded Semantics' registry-first, non-interchangeable architecture. First, the system must not use vectors for identity. Embeddings must remain strictly at the edges of the architecture, functioning as candidate retrieval mechanisms (Expression \-\> Vector \-\> Approximate Search \-\> Candidate Concepts). They are too unstable, anisotropic, and lossy to define the core semantic identity. Second, the system's vector-matching engine must discard pure cosine similarity in favor of Cross-Domain Similarity Local Scaling (CSLS) to mitigate the massive hubness problem prevalent in multilingual spaces40. Third, when an expression is linked to a Concept, the connection must include a required "Residue Payload." If translating Javanese to English, the payload \[REGISTER \= Krama Inggil\] is passed along. If the target language cannot express this grammatically, the system can output a fallback (e.g., adding "sir/ma'am") or explicitly invoke an abstention, preventing semantic corruption. Finally, the registry should anchor Concepts using a Medoid approach (the most central, actually existing expression vector in the cluster) rather than a mathematical centroid, avoiding the empty-space and polysemy problems.

20. Website Content Recommendations

To effectively communicate this philosophy to users, researchers, and potential integrators, EmbeddedSemantics.com should feature content that challenges prevailing industry assumptions:

• The "Translation is Lossy" Manifesto: A clear, accessible explanation of why replacing everything with English erases human culture and grammar, outlining the registry-first approach.
• The Anatomy of a Concept: A visual breakdown showing how a single Concept (e.g., GREETING) branches into dozens of language-specific expressions, each with distinct Semantic Residue tags representing register, locale, and evidentiality.
• Vector Geometry Explainer: A simplified guide on why AI models and embeddings get confused by "hub" words and polysemy, proving why human-curated registries and discrete semantic graphs are necessary companions to neural models.

21. Proposed Deep-Dive Articles

To establish thought leadership, the following deep-dive articles should be published:

1. The Hubness Trap: Why Your Multilingual AI Thinks Everything is Related. (Exploring anisotropy, Top-PC removal, and why CSLS is mandatory for cross-lingual retrieval).

2. Evidentiality and AI: When Algorithms Erase How We Know What We Know. (Exploring Tuyuca, Turkish, and the catastrophic loss of epistemological grammar in modern LLMs).

3. Beyond Translation: Building the World's First True Semantic Registry. (Explaining the tradeoff between continuous vector spaces and discrete relational graphs for modeling semantic residue).

22. Proposed Interactive Demonstrations

• The Residue Calculator: A user inputs a culturally specific word or grammatically complex phrase (e.g., Schadenfreude or a Tuyuca evidential verb). The demo shows the primary concept mapping and dynamically displays the "Semantic Residue" tags being stripped away if it were forced through standard machine translation, highlighting what is lost.
• The Hubness Explorer: A 3D interactive scatter plot of a multilingual embedding space. The user can click a "Centroid Average" button and watch the point fly into a dense cluster of unrelated words (the hub), visually demonstrating why medoids and CSLS are required for stable prototype generation.

23. Diagrams and Visualizations

1. The Architecture Pipeline: A flowchart detailing the path: \[Source Expression\] \-\> \[Embedding Evidence\] \-\> \[CSLS Concept Matching\] \-\> \[Primary Concept ID\] \-\> \[Residue Extraction\] \-\> \[Target Language Generation with Compensatory Logic or Abstention\].

2. Anisotropy Cone: A 3D cone visualization showing how Transformer embeddings do not occupy a sphere uniformly, leading to false positive semantic matches and centroid drift.

3. Residue Taxonomy Graph: A node-link diagram showing the Primary Concept connected via discrete edges to Attributes (Tense, Register, Evidentiality) and Relational Edges, illustrating the hybrid taxonomy model.

24. Open Research Questions

The implementation of this architecture leaves several critical avenues for future research:

1. Dynamic Residue Generation: Can Large Language Models (LLMs) be prompted to automatically and reliably extract the exact Semantic Residue (using the formal taxonomy) by comparing a source expression to its primary concept prototype, or will this always require human-in-the-loop curation?

2. Cross-Lingual Polysemy: How can the architecture automatically detect when an expression spans two distinct concepts in the registry (sense clustering) without relying entirely on manual human auditing?

3. Compositional Limits: At what point does a lexical gap become so vast that mapping it to a single primary concept is fundamentally incorrect, requiring a completely new, compound concept node to prevent unacceptable semantic distortion?

25. Annotated Bibliography

This research is founded upon a synthesis of computational geometry, natural language processing, and comparative linguistics. The following core texts inform the architectural recommendations:

1. Sakhawat, A., et al. (2026). Hubness, Not Anisotropy, Drives Cross-Lingual Retrieval Asymmetry in Multilingual Embedding Models.

\[cite: 40, 42\]

• Relevance: This paper fundamentally alters the understanding of vector-space pathologies. It proves that hubness—not anisotropy—dominates retrieval failures in multilingual spaces, validating the architectural requirement for Cross-Domain Similarity Local Scaling (CSLS) over pure cosine similarity when interacting with concept prototypes.

2. Reimers, N., & Gurevych, I. (2020). Making Monolingual Sentence Embeddings Multilingual using Knowledge Distillation.

\[cite: 6, 7, 50\]

• Relevance: Demonstrates how vector spaces are aligned through translation via student-teacher distillation. It explicitly shows the mechanism by which semantic residue is currently erased in modern NLP, as the student model is forced to adopt the semantic blind spots of the English teacher model.

3. Barnes, J. (1984). Evidentials in the Tuyuca Verb.

\[cite: 25, 26, 27\]

• Relevance: Crucial documentation of grammatical evidentiality. It proves that grammar-carried meaning (epistemological stance) is a mandatory component of truth-conditional semantics in many languages, demonstrating that English-centric representations cause catastrophic meaning loss.

4. Aikhenvald, A. Y. (2000). Classifiers: A Typology of Noun Categorization Devices.

\[cite: 31, 51\]

• Relevance: A foundational text on numeral and shape classifiers (such as those in Tzeltal and Mandarin). It provides the theoretical basis for modeling the SHAPE\_CLASSIFIER residue attribute, demonstrating that counting is intimately tied to physical ontology in non-European languages.

5. Chen, J., et al. (2024). BGE M3-Embedding: Multi-Lingual, Multi-Functionality, Multi-Granularity Text Embeddings.

\[cite: 4, 52\]

• Relevance: Details the current state-of-the-art in multi-granularity embedding systems, illustrating both the unprecedented retrieval capabilities and the persistent limits of raw vector representations for defining true cross-lingual semantic identity.

Works cited

1. Do Language Models Understand Honorific Systems in Javanese? \- Monash University, https://research.monash.edu/en/publications/do-language-models-understand-honorific-systems-in-javanese/

2. Construction of a Dialect-Sensitive Javanese Semantic Lexicon to Support Machine Translation Systems, https://jurnal.polibatam.ac.id/index.php/JAIC/article/download/13379/3832/44240

3. Resolving Linguistic Asymmetry: Forging Symmetric Multilingual Embeddings Through Asymmetric Contrastive and Curriculum Learning \- MDPI, https://www.mdpi.com/2073-8994/17/9/1386

4. Language-agnostic BERT Sentence Embedding | Request PDF \- ResearchGate, https://www.researchgate.net/publication/361056916\_Language-agnostic\_BERT\_Sentence\_Embedding

5. Making Monolingual Sentence Embeddings Multilingual using Knowledge Distillation | Request PDF \- ResearchGate, https://www.researchgate.net/publication/347234503\_Making\_Monolingual\_Sentence\_Embeddings\_Multilingual\_using\_Knowledge\_Distillation

6. Multilingual Sentence Embeddings via Distillation | PDF \- Scribd, https://www.scribd.com/document/478399947/2004-09813v1-pdf

7. Making Monolingual Sentence Embeddings Multilingual using Knowledge Distillation \- arXiv, https://arxiv.org/abs/2004.09813

8. Language Agnostic Multilingual Sentence Embedding Models for Semantic Search | Primer, https://primer.ai/blog/language-agnostic-multilingual-sentence-embedding-models-for-semantic-search

9. University of Indonesia at SemEval-2025 Task 11: Evaluating State-of-the-Art Encoders for Multi-Label Emotion Detection \- ACL Anthology, https://aclanthology.org/2025.semeval-1.279.pdf

10. arXiv:2504.12879v1 \[cs.IR\] 17 Apr 2025, https://arxiv.org/pdf/2504.12879

11. Benchmarking and Building Zero-Shot Hindi Retrieval Model with Hindi-BEIR and NLLB-E5 \- ACL Anthology, https://aclanthology.org/2025.naacl-long.220.pdf

12. Book: Universal Networking Language (Computational Linguistics, Translation), http://www.cicling.org/2005/UNL-book/

13. Universal Networking Language \- Wikipedia, https://en.wikipedia.org/wiki/Universal\_Networking\_Language

14. UNIVERSAL NETWORKING LANGUAGE (UNL), https://language.worldofcomputing.net/unl/universal-networking-language-unl.html

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

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

17. Cultural transmission and semantic loss in the English translation of V. J. James's Nireeswaran \- Dialnet, https://dialnet.unirioja.es/descarga/articulo/10174094.pdf

18. 5 From untranslatability to translatability \- Amazon S3, https://s3-eu-west-1.amazonaws.com/s3-euw1-ap-pe-ws4-cws-documents.ri-prod/translationstudies/sample-docs/9781408289839-sample.pdf

19. Strong Linguistic Relativity: A Continental Sense of Language and Being, https://digitalcommons.lmu.edu/cgi/viewcontent.cgi?article=1551\&context=honors-thesis

20. THE LIMITS OF MY LANGUAGE, https://ftp.rudraksham.com/story/0TW/026/2SVz6Y/the\_limits\_of\_my-language

21. 28 Untranslatable Words from Around the World, https://ourworldenglish.com/28-untranslatable-words-from-around-the-world/

22. Eleven Untranslatable Words From Other Cultures | by Maptia \- Medium, https://medium.com/@Maptia/eleven-untranslatable-words-from-other-cultures-94ec1b9f5741

23. 203 Most Beautiful Untranslatable Words \[The Ultimate List: A-Z\] \- The Intrepid Guide, https://www.theintrepidguide.com/untranslatable-words-ultimate-list/

24. Evidentials and source knowledge in Turkish \- PubMed, https://pubmed.ncbi.nlm.nih.gov/19787641/

25. Seeing and knowing \- John Benjamins, https://www.jbe-platform.com/content/journals/10.1075/fol.22006.ber?crawler=true

26. Full article: The development of the concept of ʽevidentialityʼ and its exogenous application to European languages \- Taylor & Francis, https://www.tandfonline.com/doi/full/10.1080/17597536.2023.2171701

27. Chapter Semantic Distinctions of Evidentiality \- WALS Online, https://wals.info/chapter/77

28. Four types of evidentiality in the native languages of Brazil \- ResearchGate, https://www.researchgate.net/publication/276157738\_Four\_types\_of\_evidentiality\_in\_the\_native\_languages\_of\_Brazil

29. A Reconsidered Language Model Approach to the Problem of Verb Aspectual Pairs in the Russian Language, https://jle.hse.ru/article/view/23887

30. systems of nominal \- classification, http://www.ddl.cnrs.fr/Colloques/3L\_2008/3LCourseMaterial/Maya/Zavala\_2000\_Multiple\_Classifier\_Systems\_Akatek.pdf

31. Tracking semantic relatedness: numeral classifiers guide gaze to visual world objects, https://www.frontiersin.org/journals/language-sciences/articles/10.3389/flang.2023.1222982/full

32. Chapter Numeral Classifiers \- WALS Online, https://wals.info/chapter/55

33. The conceptual building blocks of kinship terminologies \- Griffith Research Online, https://research-repository.griffith.edu.au/items/51591175-3c18-41e8-b12a-a2a12d8b28d7

34. Homo Symbolicus: Ernst Cassirer, the Philosopher Who Won the Argument and Lost the Century \- Medium, https://medium.com/@krigerbruce/homo-symbolicus-ernst-cassirer-the-philosopher-who-won-the-argument-and-lost-the-century-02234ff4c7b9

35. Lexico-Semantic Aspects of Complex Words (Part I), https://www.cambridge.org/core/books/complex-words/lexicosemantic-aspects-of-complex-words/372EB6685D509FA7CE76706BE5C4ADC7

36. Complex Words: Advances in Morphology 9781108490290, 9781108780643, 9781108748377 \- DOKUMEN.PUB, https://dokumen.pub/complex-words-advances-in-morphology-9781108490290-9781108780643-9781108748377.html

37. Classification of Forensic Autopsy Reports through Conceptual Graph-based Document Representation Model | Request PDF \- ResearchGate, https://www.researchgate.net/publication/324974336\_Classification\_of\_Forensic\_Autopsy\_Reports\_through\_Conceptual\_Graph-based\_Document\_Representation\_Model

38. Event knowledge and models of logical metonymy interpretation, https://d-nb.info/1058859552/34

39. Anisotropy in Embedding Representations \- Emergent Mind, https://www.emergentmind.com/topics/anisotropy-in-embedding-representations

40. \[2605.26575\] Hubness, Not Anisotropy, Drives Cross-Lingual Retrieval Asymmetry in Multilingual Embedding Models \- arXiv, https://arxiv.org/abs/2605.26575

41. (PDF) An Isotropy Analysis in the Multilingual BERT Embedding Space \- ResearchGate, https://www.researchgate.net/publication/361063706\_An\_Isotropy\_Analysis\_in\_the\_Multilingual\_BERT\_Embedding\_Space

42. Hubness, Not Anisotropy, Drives Cross-Lingual Retrieval Asymmetry in Multilingual Embedding Models \- arXiv, https://arxiv.org/html/2605.26575v1

43. Word translation without parallel data \- arXiv, https://arxiv.org/pdf/1710.04087

44. Loss in Translation: Learning Bilingual Word Mapping with a Retrieval Criterion \- ACL Anthology, https://aclanthology.org/D18-1330.pdf

45. An Isotropy Analysis in the Multilingual BERT Embedding Space \- ACL Anthology, https://aclanthology.org/2022.findings-acl.103.pdf

46. A comprehensive analysis of the parameters in the creation and comparison of feature vectors in distributional semantic models f, https://doktori.bibl.u-szeged.hu/10120/1/AndrasDoboThesis2019.pdf

47. Hierarchical Mapping for Crosslingual Word Embedding Alignment | Transactions of the Association for Computational Linguistics \- MIT Press Direct, https://direct.mit.edu/tacl/article/doi/10.1162/tacl\_a\_00320/96459/Hierarchical-Mapping-for-Crosslingual-Word

48. Benchmarking and Building Zero-Shot Hindi Retrieval Model with Hindi-BEIR and NLLB-E5, https://arxiv.org/html/2409.05401v3

49. \[Literature Review\] Hubness, Not Anisotropy, Drives Cross-Lingual Retrieval Asymmetry in Multilingual Embedding Models \- Moonlight, https://www.themoonlight.io/en/review/hubness-not-anisotropy-drives-cross-lingual-retrieval-asymmetry-in-multilingual-embedding-models

50. Making Monolingual Sentence Embeddings Multilingual Using Knowledge Distillation, https://www.semanticscholar.org/paper/Making-Monolingual-Sentence-Embeddings-Multilingual-Reimers-Gurevych/8f34ee2ec88e8b19b2736de55eb170539d26e527

51. The semantic domains of numeral classifiers in Kammu \- Lund University Publications, https://lup.lub.lu.se/luur/download?func=downloadFile\&recordOId=1323508\&fileOId=1323509

52. Claim2Vec: Embedding Fact-Check Claims for Multilingual Similarity and Clustering \- arXiv, https://arxiv.org/html/2604.09812v1