Lingotran Engine

A scalable workflow for extracting structured linguistic data from various language-learning PDFs, enabling the automated conversion of textbook content into structured JSON formats for personal study tools.

Supported Languages

• German — shipped: 636 pages, fully transcribed, QA-verified, exported (german/extracted/_exports/).
• French — in progress: 3 of 5 new PDFs done via the shared _engine/ pipeline (Cosmopolite A1 Méthode 224 pages, Tricolore 1 180 pages, Tricolore 2 180 pages), 2 remain (Saison 2, Cosmopolite 5); plus 2 legacy books on an older, separate pipeline (conjugaison-a1-a2 partially transcribed, revision-2 not started).
• Japanese, Portuguese, Romanian, Russian, Spanish — not yet started (empty stub folders).

Overview

This engine is designed to parse multilingual text and structures from learning materials and translate them into a unified format for data consumption in learning applications.

Cross-language export conventions

These apply to every language's pipeline, whether it uses the shared _engine/ or its own bespoke _tools/ (e.g. German's), so exports stay usable to anyone regardless of which source language they were extracted from:

• All taxonomy/enum column values are English, always — section, activity_type, content_type, topic, word_class, status, qa, level (item_type too). These are our own categorization scheme, not text copied from the book, so they must read in English no matter what language the source material is in. (Found and fixed 2026-07-28: German's section column used German words — hoeren/lesen/sprechen/ schreiben — instead of listening/reading/speaking/writing, across all non-frozen German books; each language's agent_enrich.md now states this explicitly.)
• All column names are English, always, and identical across languages — the header row is part of the schema, not the source material. (Found and fixed 2026-08-10: German's questions sheet named its exercise-group column teil where French named the same concept part, so the two languages' sheets did not line up for anyone reading both. Renamed to part; German's questions sheet also gained the level column French already had, so both languages now emit a byte-identical header.) A column that exists in one language and genuinely has no source data in another — chapter, which German books do not carry — is simply absent there rather than shipped permanently empty.
• Verbatim/quoted fields stay in the source language — the question text itself, answer options, examples, and printed part/unit labels — these are direct quotes from the book and must not be translated or reworded, unlike the taxonomy columns above. Note the distinction from the rule above: the part column name is English, while the values in it stay exactly as printed ("Teil 1", "Übung 2", "Unité 3").
• Before adding a new language's pipeline, re-check its agent_enrich.md/equivalent enrichment prompt for this rule, and spot-check a sample export for stray non-English taxonomy values before considering that language's export "done."