Automatic Generation of Dyslexia-Friendly French Educational Texts Using Fine-Tuned LLMs: A Phonics-Aware Approach for Moroccan Primary School Children

Abstract

Dyslexia is one of the most common learning disabilities among school-age children, affecting an estimated 10 to 15% of the global school population. In the Moroccan context, where French serves as the primary language of instruction in many private primary schools, children with reading difficulties face a double challenge: they must decode a language that is not their mother tongue, using standard textbooks that were never designed with accessibility in mind. We describe an end-to-end system for automatically generating dyslexia-friendly versions of French primary school texts, targeting children aged 6 to 8 (CP and CE1). The approach relies on fine-tuning Mistral-7B-Instruct-v0.2 using LoRA on a manually constructed dataset of 362 text pairs, each simplified according to a 13-rule dyslexia-friendly referential. Three models were evaluated on a fixed test set of 36 pairs: Model 1 (99 pairs, BLEU 61.53, SARI 79.50), Model 2 (243 pairs, BLEU 68.29, SARI 83.09), and Model 3 (294 pairs, BLEU 70.02, SARI 84.64). Model 3 substantially outperforms both zero-shot Mistral (BLEU 26.85, SARI 53.01) and the best published French text accessibility system, ETR-fr/Mistral+LoRA (SARI 42.27), while using 44% fewer training pairs than ETR-fr. Readability analysis of model outputs confirms that Model 3 produces text measurably more readable than the original on all five metrics. To our knowledge, this work presents the first trainable, pipeline-integrable system for French graphemic syllabification aligned with CP/CE1 phonics instruction, achieving approximately 93% segmentation accuracy on unseen texts — a meaningful first result given the complete absence of production-ready French phonological syllabification tools. The fine-tuned model is integrated into LectureFacile, a web application (https://huggingface.co/spaces/lsadouk1111/lecturefacile-dyslexie) that transforms any photographed textbook page into an accessible version in real time.