Text Simplification for Children: Evaluating LLMs vis-à-vis Human Experts
Anastasia Smirnova, Kyu beom Chun, Wil Louis Rothman, Siyona Sarma
Abstract
Large Language Models (LLMs) can facilitate teaching and learning by generating educational content for specific grade levels.Generation of age and grade-appropriate materials often involves text simplification.Previous work evaluating LLMs' text simplification for children in educational context showed that LLMs can reduce lexical complexity of a text and improve its readability, but they often use vocabulary that is still too difficult for the targeted grade.In this study we focus on GPT-4o, the most advanced LLM at the time of writing.We evaluate its ability to simplify text for elementary school children and compare its performance vis--vis the human baseline.We show that GPT-4o can successfully reduce text complexity on lexical, syntactic, and discourse levels.However, compared to human experts, it generates syntactically more complex text.This suggests that further fine-tuning with focus on syntax is needed before LLM-generated output reaches the intended user group.