LLMs are probabilistic. Language is structured. Suzon bridges both worlds.
Suzon is an explainable linguistic layer that sits before and after large language models — normalizing language on the way in, analyzing structural consistency on the way out. It is inspired by the pioneering work of Belgian computational linguist Jacques Blois, whose archive you'll find further down this page.
Suzon operates as a linguistic layer around large language models. Before generation, it analyzes morphology and normalizes linguistic forms into a structured representation. After generation, it evaluates structural consistency and exposes the linguistic reasoning behind the response.
We call it a linguistic immune system: the way an immune system detects pathogens before they spread, Suzon detects linguistic inconsistencies before and after a language model produces an answer.
The goal is threefold: explainability (every transformation can be traced), measurable bias reduction (structured input reduces ambiguity a model might otherwise resolve arbitrarily), and multilingual precision (morphological analysis handles conjugations, declensions, and compounds across languages).
Suzon does not replace the language model — it wraps it. Deterministic linguistic analysis handles what is structured about language; the LLM handles what is generative. Each step below is inspectable.
The core method — decompose a word into roots and affixes, resolve it against a structured system, and only then let the machine act on it — was worked out in the early 1960s, in Brussels, by a linguist named Jacques Blois. Suzon draws inspiration from the methodological principles he pioneered.
What follows is the archive: five rooms documenting where that method came from, every claim backed by primary or peer-reviewed sources.
In 1961, a small team at the Université libre de Bruxelles began work on DICAUTOM (Dictionnaire Automatique) — a system designed to automate dictionary lookup for human translators, rather than replace them outright.
Blois joined this effort with a specific problem to solve: French is a heavily inflected language, and a translator's dictionary needs to recognize a word however it appears — conjugated, pluralized, or combined with prefixes and suffixes — and trace it back to a single entry.
Source: Université libre de Bruxelles, activity report of the Groupe de Linguistique Automatique, 1963. Read the report →
Blois and E. Morlet published Morphologie du français pour la traduction automatique — a method for decomposing French words into roots and affixes, so a machine could reduce any inflected form to a single dictionary headword before looking it up.
It's a modest-sounding idea that solves a real problem: without it, a dictionary lookup system would need a separate entry for every conjugated, pluralized, or derived form of every word — an impossible task at the scale of a living language.
Source: J. Blois & E. Morlet, Morphologie du français pour la traduction automatique, Publications Office of the European Union, 1962. Read the publication →
Two decades later, at the Institut supérieur de traducteurs et interprètes (ISTI), Blois moved from morphological theory to applied databases. Working on IBM's STAIRS system, he built two tools: CLEF, a lookup database for translated technical terms, and LEXI, a bibliographic and lexical database built from student research on neologisms — with automatic lemmatization, so a single query could retrieve every inflected or compound form of a term.
This terminal is a stylistic homage to the STAIRS interface Blois described in his 1983
article — not a reproduction of its data. For the original search transcripts (query
display, query agent de dissuasion), see the source below.
Source: Jacques Blois, "Les banques de données de l'I.S.T.I.," Équivalences, vol. 14, no. 2–3, 1983, pp. 61–66. Read on Persée →
DICAUTOM is discussed as one of the direct forerunners of Eurodicautom, the European Commission's terminology database, launched in 1975. Eurodicautom was folded into IATE in 2007 — the terminology base still used by EU institutions today.
Source: Gwénaël Glâtre, "Le traducteur et l'ingénieur « Eurodicautom », années 1960," Encyclopédie d'histoire numérique de l'Europe, Sorbonne University. Read the entry →
Every historical statement on this site is backed by primary or peer-reviewed sources.
The Jacques Blois Research Initiative was founded in 2026, building upon the pioneering work Jacques Blois began in Brussels in 1961. The archive preserves that work; Suzon applies it. Join us in building explainable, multilingual AI.