Scholarlink

<p>The complexity of text significantly impacts reading comprehension, especially for young readers. Identifying complex vocabulary and then replacing it with a simple alternative is essential for improving reading accessibility among children. This study investigates the effectiveness of large language models (LLMs) to perform complex word identification (CWI) within children’s stories through a zero-shot prompting approach. We designed a child-centric annotation interface to create ground truth by learners aged 8–10, capturing judgments at the document level rather than isolated sentences. To ensure explainability, we conducted statistical analyses of linguistic features across complex and simple words, revealing how structural features impact model decisions. We evaluate four LLMs, GPT-4o and open-source models Llama3, Mistral, and Gemma2, against these annotations. Results show that GPT-4o demonstrates better alignment with children’s perception of word complexity, achieving a balanced trade-off between recall and precision. These findings establish the effectiveness of LLMs and emphasize the value of considering target user perception to evaluate the model’s decisions.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Through child’s eyes: evaluating LLMs for complex word identification in children’s stories

  • 期刊论文

  • 发表日期：

    2025年12月8日

查看全文
  • Tamma RajyaLakshmi,
  • K. S. Kuppusamy

摘要

The complexity of text significantly impacts reading comprehension, especially for young readers. Identifying complex vocabulary and then replacing it with a simple alternative is essential for improving reading accessibility among children. This study investigates the effectiveness of large language models (LLMs) to perform complex word identification (CWI) within children’s stories through a zero-shot prompting approach. We designed a child-centric annotation interface to create ground truth by learners aged 8–10, capturing judgments at the document level rather than isolated sentences. To ensure explainability, we conducted statistical analyses of linguistic features across complex and simple words, revealing how structural features impact model decisions. We evaluate four LLMs, GPT-4o and open-source models Llama3, Mistral, and Gemma2, against these annotations. Results show that GPT-4o demonstrates better alignment with children’s perception of word complexity, achieving a balanced trade-off between recall and precision. These findings establish the effectiveness of LLMs and emphasize the value of considering target user perception to evaluate the model’s decisions.