The Bridge Method

Status: decided design — the canonical Bridge Method adaptation (immersive#280); current.

The canonical definition of Immersive’s Bridge Method — how the EuRom5 intercomprehension method (reviewed in detail in the EuRom5 source in the research knowledge base - admin: Knowledge Base > Sources, SLA research domain; #407, moved there in #540) is adapted into Immersive’s product. The design doc (immersive_redesign_extension_design.md, “Product North Star”) states why this is the north star; this document defines what the method is in Immersive terms and how each surface applies it.

Definition

A learner who already knows one or more of the six Bridge languages — English, Spanish, French, Italian, Portuguese, Catalan — acquires another one faster when the product makes explicit, structural use of what they already know: cognates, shared constructions, parallel morphology, and transferable reading strategies, instead of teaching the target language in isolation.

Canonical language order

Any multi-language display orders the languages PT – ES – CA – IT – FR – EN (decision recorded in #407). The rationale, from EuRom5: the Romance five are always presented along the geographic continuum — adjacent columns are the most similar pairs (PT↔ES, ES↔CA, CA↔IT), with French the outlier at the edge. Reading across a row, each cell is then a small step from its neighbour, so the eye tracks the gradual transformation of a single form: pão – pan – pa – pane – pain. Alphabetical or arbitrary ordering would interleave distant languages and destroy that gradient, which is precisely what teaches. English is not on the Romance continuum; it sits at the French edge because its Latinate vocabulary layer arrived mostly via French, making FR→EN the smoothest final step.

UI usually shows only the learner’s known languages plus the target, but whatever subset appears keeps this relative order.

What we adopt from EuRom5, what we add, what we don’t copy

Adopted (per-finding source detail lives on the EuRom5 source in the research knowledge base (admin: Knowledge Base > Sources, SLA research domain; #540)):

Known languages are a resource, not interference — surfaced, not suppressed.
Global-first reading with the right to approximation: the reader succeeds by understanding the text, not by translating every word.
The aid taxonomy below, including aids whose message is “you don’t need help here.”
Relative opacity: aids are filtered by what is opaque to this learner, given their actual known languages.
Scaffolding fade: aid density decreases as demonstrated level rises.
Receptive competence as a legitimate, separately-tracked axis (CEFR partial competences): comprehension may legitimately run years ahead of production.
Audio as a comprehension aid while reading (prosody segments syntax; sound can be transparent where spelling is opaque).
A contrastive “grammar of reading” layer selected by frequency × observed difficulty, cross-referenced from texts.

Added by Immersive (not in EuRom5):

English as a sixth Bridge language (Latinate/French-derived vocabulary layer).
Spaced repetition (FSRS): comprehension moments become durable review items.
Production practice: writing, correction, dictation, speaking modes.
Personalization: EuRom5 prints one superset of aids and asks teachers to filter; Immersive computes the filter per user.

Not copied: EuRom5’s texts and printed aids are copyrighted (© Hoepli, press articles). Immersive ships only its own or licence-clean sourced texts — and per project decision (#407), texts are never LLM-generated; only aids are.

The two approaches

Issue #407 named two approaches; both are supported, in order:

Approach A — target focus with bridge hints (default, first). The learner studies one target language; every surface annotates against their known Bridge languages where it speeds comprehension. This is a delta on the existing product: the four shipped bridge components plus the aid taxonomy below, filtered by user_languages.
Approach B — intercomprehension reading mode (later, Documents-scoped). EuRom5’s actual engine is alternation: several languages per session, switching every ~30 minutes, so that comprehension strategies (not vocabulary) transfer and become language-independent. As a future Documents mode: a reading session that alternates texts across the learner’s chosen language set with the same aid apparatus. Captured as a draft issue; built only after A proves out.

Aid taxonomy

EuRom5’s apparatus is a spectrum from “don’t translate” to “translate”, ordered here roughly by how much they give away. Every aid is keyed to a line/token (the annotated line-numbered reading component carries this role).

Aid	EuRom5 original	Immersive form
Title orientation	Title pre-translated into the other four languages before reading	Display-language title gloss/theme cue shown before the reader opens
Skip marker	Portion marked “effacée” — inferable once context is understood	Mark a span “safely skippable”; reader is told guessing is enough
Phantom/inference nudge	Empty-box symbol: meaning inferable from immediate context or collocation — don’t translate yet	Progressive disclosure: first reveal step is “guessable from context — try it” with the translation behind a second step
Listen hint	Word-level headphone icon: the sound reveals what the spelling hides	Per-word “listen” affordance as an aid class distinct from translation
Already-seen pointer	Arrow + line number: same/derived form appeared earlier	Link back to the earlier occurrence in this document
Morphology decomposition	storielle = storia + -elle; panaderos via PAN	Word-mode hint splitting a token into recognizable parts, mapped to known-language cognates
Structure skeleton	S/V/O reduction, `[...]` for interruptions, `[brackets]` for reconstructed elements	The shipped sentence-decomposition component, extended with skeleton-first reveal for hard sentences
Alignment box	One construction vertically aligned across the five languages	The shipped grammar-comparison callout, restricted to the learner’s known languages + target
Lexical equivalence	Line-keyed table row across all five languages	The shipped equivalence-table component; two registers below
Grammar cross-ref	[§n] into the grammar of reading	Link from a reader annotation to the relevant contrastive grammar note

Two registers for the translation-type aids, both deliberately not professional translation (EuRom5 §3.5):

Calque (“par le bas”): word-shape-preserving gloss that exposes the transparency between languages.
Sense gloss (“par le haut”): context-dependent meaning arrived at by inference.

The reader’s translation reveal should be able to show both.

Core principles for implementation

Relative opacity. An aid is only an aid if the item is opaque to this learner. ES llegados needs nothing for an Italian-knower (arrivati) and a full aid for an English-only knower. Aid computation takes the user’s user_languages (all known Bridge languages with levels) as input, not a single fluent language. This is the prerequisite the bridge-language data-model issue covers.
Aids are data, not ephemeral output. Aids are LLM-drafted initially but stored in the database, cached per (document, known-language-set), and updatable — by humans or by LLM passes driven by user feedback. This mirrors how EuRom5’s annotations were built: from learner-reported difficulties collected over years, not from linguists’ predictions.
Scaffolding fades. Aid density is a function of the learner’s demonstrated receptive level (EuRom5: dense in tranche A, sparse in tranche C). New readers get the full apparatus; advanced readers get lexical tables and grammar refs only.
Never require exhaustive translation. The reader’s success state is global comprehension. No mechanic should push word-by-word completionism; the approximation right is a feature.
Don’t over-warn on false friends. Experimentation found more “vrais amis” than “faux amis”; flag only pairs that actually break comprehension.
Default aid prominence follows observed learner interest: lexicon first, then syntax, then morphology.
Receptive and productive progress are separate axes. A learner may read at B2 and produce at A2; Profile should be able to show both (design note in #298).

Application per surface

Documents (primary home): the full aid taxonomy in the reader (comprehension-aid layer issue); synchronized listen-while-reading; title orientation; graded reading library from licence-clean sourced texts with EuRom5-style length tranches and fading aids; later, the intercomprehension reading mode (approach B).
Grammar: a contrastive grammar-of-reading layer — receptive-focused comparison notes selected by frequency × difficulty, targeted by cross-refs from reader annotations; the existing 1,913 grammar notes reviewed for Bridge Method fit (impact-review issue). The shipped grammar-comparison callout is the display primitive.
Verbs: cross-language conjugation comparison — the same tense/person aligned across the six languages (EuRom5’s -avam/-aban/-aven/-avano/-aient vertical alignment; conjugation data already exists for all six).
Decks/cards: contrastive card templates, including inference cards (document.word.inference.v1: sentence context with a masked word; the learner infers before revealing) generated from phantom-word aid interactions.
Profile: user_languages capture (which Bridge languages, at what level A1–C2/native); receptive vs productive axes; bridge-aware recommendations.
Chat: context includes the learner’s known languages so explanations can bridge (“like appena in Italian, which you know”).

Language model

The user schema this method requires:

target_language — the language being learned.
fluent_language — the language aids, translations, and explanations are presented in. One value per user, set at sign-up and editable on the profile.
user_languages — one row per (user, Bridge language) with level a1…c2 or native. This is the knowledge model the aid engine reads, and the reason multi-language knowledge is representable at all: fluent_language can only ever hold one answer, and a learner who knows English and French needs both on record before relative opacity can be computed for them.

Positioning evidence

For marketing/positioning claims (see immersive_business_plan.md): EuRom5 is an EU- and French-government-funded line of research (1990 experiments → EuRom4 1997 → EuRom5 2010) whose published result is autonomous newspaper reading across four related languages in about 40 guided hours (~B1–B2 receptive), grounded in the CEFR’s partial-competences framework. Honest caveat whenever cited: that evidence covers comprehension only, among Romance-L1 adults — Immersive’s extensions (English in the set, production practice, spaced repetition) go beyond what EuRom5 demonstrated.