Marsji: A Concept-ID Semantic Layer for Cross-Language Lexical Mapping and Accessible Interfaces

# Marsji: A Concept-ID Semantic Layer for Cross-Language Lexical Mapping and Accessible Interfaces **A Proposal and Status Report** **Author:** Gino Rouzee **Affiliation:** Independent researcher **First version:** May 13, 2026 (Marsji Day) **This version:** November 2026 **Repository:** GitHub/Marsji-Language-Project **Web interface:** marsji.com **10.5281/zenodo.23857698 --- ## Abstract This paper describes Marsji, a semantic layer that assigns a fixed identifier (ID) to each concept and links that ID to words in many natural languages. Instead of mapping words directly from one language to another, every word is mapped to a shared concept ID, and from there to its counterparts in other languages. The current dataset contains approximately 2.084 million word-to-concept relationships across 50 languages, including a core vocabulary of roughly 1,000 structural words per language. The paper presents the three-layer architecture of the system, the design of its ID scheme, the current state of the dataset including its known quality limitations, and what the approach can and cannot do. It also proposes, as a direction for future research, a learnable tactile and visual code for concept IDs, inspired by the six-dot braille cell, intended as an additional communication aid for blind, deaf and deafblind users. The claims in this paper are limited to what the current implementation demonstrates; the accessibility application is presented as a hypothesis to be tested together with the communities concerned. ## 1. Introduction Most translation technology today relies on statistical or neural models that learn correspondences between languages from large amounts of text. These systems are powerful, but their output is not deterministic: the same input can produce different translations, and it is difficult to trace why a particular word was chosen. Marsji explores a complementary approach. Its starting point is the observation that many concepts, such as *love*, *water* or *mother*, are shared across languages even though the words for them differ. If each such concept receives a stable ID, then words in different languages can be linked through that ID in a transparent and reproducible way. A lookup from Urdu to Spanish, for example, becomes: Urdu word → concept ID → Spanish word. The guiding idea behind the project can be summarized as: *the power of information lies not in the number of symbols, but in the relationships between them.* Marsji is an attempt to make those relationships explicit, stable and machine-readable. Marsji is not intended to replace natural languages, nor to replace machine translation. It is a structured lexical layer that can support search, indexing, vocabulary learning and, potentially, accessibility tools. ## 2. Related Work The idea of a shared conceptual layer between languages has a long history, and Marsji builds on several existing approaches: - **Universal Networking Language (UNL)**, developed under the United Nations University from the late 1990s, used "Universal Words" as language-independent concept identifiers, with the aim of serving as an interlingua for translation. - **WordNet** (Miller, 1995) organizes English words into sets of synonyms representing distinct senses, linked by semantic relations. Many WordNets for other languages have since been linked to it. - **BabelNet** (Navigli & Ponzetto, 2012) combines WordNet and Wikipedia into a large multilingual semantic network in which concepts carry identifiers across hundreds of languages. - **Wikidata lexemes** link words in many languages to language-independent items. - **Blissymbolics** (Bliss, 1965) is a system of symbols representing concepts rather than sounds, used today in augmentative and alternative communication. Marsji differs from these mainly in three respects: it introduces its own compact word forms for each concept (the Marsji core vocabulary), it adds a deterministic reversible encoding layer, and it explores a direct link between concept IDs and a tactile/visual code. A systematic comparison with BabelNet in particular is part of the planned evaluation (Section 8). ## 3. System Architecture Marsji consists of three layers. ### 3.1 Surface layer (natural language) The words of natural languages as they are written: *love* (English), *amour* (French), *محبت* (Urdu). Words with several meanings (homonyms and polysemous words) are stored as separate entries, one per sense, so that each entry can be linked to exactly one concept. ### 3.2 Matrix layer (reversible encoding) Each surface entry is converted into a deterministic machine string using a reversible transformation: the original word can always be reconstructed exactly from its encoded form. For example, the 45-letter word *pneumonoultramicroscopicsilicovolcanoconiosis* maps to a 45-character encoded string and back without loss. This layer provides a uniform, language-independent representation for processing. It does not compress data; its purpose is consistency and exact reversibility. ### 3.3 Semantic layer (Marsji core) Each concept receives a fixed ID and a short Marsji word form, for example *nowe* for the concept of love. All surface entries that express this concept, in any language, are linked to the same ID. ## 4. The ID Scheme Every concept ID must be unique and stable over time, so that links between languages never break when the dataset grows. Marsji organizes IDs into nested blocks, with block boundaries anchored to prime numbers. This is a design choice for allocating ID ranges that can be extended without overlapping existing ranges. It should be noted that uniqueness of IDs as such can be achieved with any well-designed numbering scheme; the prime-anchored structure is one way of organizing the allocation and is not by itself a solution to ambiguity in language. The hard problem in any concept-based system is not numbering but **disambiguation**: deciding which concept a word refers to in a given context. The English word *bank* can refer to a financial institution or the side of a river. Marsji stores these as separate entries with separate IDs, but choosing the correct one in a running text still requires context, either from the user or from an additional disambiguation step. ## 5. Current Dataset | Item | Status | |---|---| | Word-to-concept relationships | approx. 2.084 million | | Languages | 50 | | Core structural vocabulary | approx. 1,000 words per language | | Specialized vocabulary | medical nomenclature (partial) | | Formats | JSON and CSV tables, expansion scripts | | Platforms | Linux, macOS, Windows (no external dependencies) | **Data quality.** A large part of the dataset was generated with the assistance of AI language models and has not yet been fully validated by native speakers. During development, several types of errors were identified, including invented abbreviations that conflict with established ones (notably in medical terminology) and outdated information presented as fact. Validation of the core vocabulary by native speakers is a priority. Users of the dataset should treat entries outside the validated core as provisional. ## 6. Capabilities and Limitations **What Marsji can do today:** - Look up a word in one language and return its counterparts in other languages via a shared concept ID - Do so deterministically: the same input always gives the same output, and every link can be inspected - Provide stable concept IDs that can be used for multilingual search and indexing - Run locally without machine-learning models at lookup time **What Marsji cannot do (yet):** - Translate full sentences. Sentences involve grammar, word order, tense, agreement, politeness forms and idioms that cannot be handled by word-to-concept mapping alone. *"It's raining cats and dogs"* contains no word that refers to heavy rain. - Disambiguate automatically. Selecting the correct sense of a word in context requires additional information or processing. - Guarantee the correctness of entries that have not yet been validated (see Section 5). ## 7. Proposed Application: A Tactile and Visual Code for Concept IDs *This section describes a research direction, not an implemented or tested system.* ### 7.1 Idea The braille cell uses six positions arranged in two columns of three: ``` (1) o o (4) (2) o o (5) (3) o o (6) ``` The proposal is to define a code in which Marsji concept IDs are expressed as sequences of such six-position patterns. The same code could be: - **felt**, through finger positions or a flat haptic plate with micro-actuators, for blind and deafblind users; - **seen**, as the corresponding hand shapes and movements, for deaf users. Because each pattern refers to a concept rather than to a word in a particular language, two people who have both learned the code could, in principle, exchange basic meanings regardless of their spoken language. A message entered as a sequence of concept codes could then be rendered in the receiver's own language through the concept ID. ### 7.2 Example scenario: a broadcast from Pakistan watched in Mexico A news anchor in Pakistan speaks Urdu. An interpreter renders the broadcast not in Pakistani Sign Language, but in the Marsji concept code: each concept is shown as its fixed hand pattern, linked to its concept ID. A deaf viewer in Mexico who has learned the same code watches the interpreter. Because the viewer associates each pattern with the same concept ID, and that ID with the Spanish word, the viewer reads the message directly in Spanish, without knowing Urdu and without translation software. The principle is the same as with numerals or traffic signs. The symbol *7* is read as *sieben*, *sept* or *siete* depending on the reader, yet everyone understands the same quantit