LLMs for Document-Level Text Simplification in Turkish Foreign Language Learning

This study presents the first investigation into the use of large language models (LLMs) for document-level text simplification targeting Turkish foreign language learning. ChatGPT-40 is utilized to simplify six Turkish stories to A1, A2, and B1 proficiency levels and evaluated on a parallel corpus ( i.e., these stories and their simplified versions at A1, A2, and B1 proficiency levels). The model is prompted with specific simplification rules and examples for these rules. We evaluate the performance using multiple metrics including BLEU, SARI, D-SARI, and BERTScore. Our results show that ChatGPT-40 can generate simplified texts comparable inlength and content to human-simplified references. This research addresses the scarcity of reading materials for Turkish learners, a challenge that many other languages also face. It demonstrates the potential of LLMs in producing level-appropriate simplified texts, opening new avenues for automated text simplification in language education.