AI can now simplify text for language learners across four languages without needing labeled training data

A new machine learning system (Re-RIGHT) can rewrite text to match a learner's proficiency level in English, Japanese, Korean, and Chinese — without requiring expensive hand-labeled pairs of original and simplified sentences. The system uses a compact 4-billion-parameter model trained on just 43,000 vocabulary examples, and outperforms larger language models like GPT and Gemini at actually matching target difficulty levels, especially for beginner learners and non-English languages.

Until now, simplifying text for language learners required either costly human annotation of parallel texts or relying on large language models that often failed at beginner levels. This system works smaller and more accurately by focusing on three specific problems: finding the right words for the target level, keeping meaning intact, and preserving sentence flow. The fact that a 4-billion-parameter model beats much larger systems on this task suggests that targeted training on real vocabulary data matters more than raw model size — which means language-learning apps could deploy this without paying for expensive API calls to commercial AI services.