ChatGPT for Text Simplification in Chinese as a Second Language Textbooks: A Comparative Study With Expert Adaptations

Abstract

ABSTRACT Based on a self‐constructed comparable corpus, this study investigates how expert Chinese language teachers and ChatGPT‐4o simplify Chinese L2 reading texts under three prompting strategies: zero‐shot, few‐shot, and instruction‐based prompting. Differences are examined across lexical, syntactic, cohesive, semantic, and content dimensions. At the linguistic level, ChatGPT‐4o‐generated texts differ significantly from expert adaptations on multiple measures; however, outputs produced with few‐shot and instruction‐based prompts approximate expert performance on several key lexical, syntactic, and cohesive indicators. Semantic features display largely convergent patterns across human and ChatGPT‐4o outputs, suggesting that semantic adjustment is strongly constrained by shared task demands. Taken together, these findings underscore the role of prompt design in linguistic simplification. At the content level, clearer divergences emerge: ChatGPT‐4o primarily relies on compression and deletion—particularly when encountering abstract or culturally embedded content—whereas expert teachers engage in pedagogically oriented adaptation, selectively reorganizing, supplementing, and re‐creating content to enhance comprehensibility and cultural accessibility. These findings point to a fundamental distinction between linguistic simplification and pedagogical content adaptation. Building on this contrast, the study proposes a three‐stage human–AI collaborative model—AI‐generated initial simplified drafts, expert refinement, and iterative feedback—aimed at integrating the efficiency of large language models with the pedagogical judgment of human teachers in Chinese as a second language material development.

Ähnliche Arbeiten