Scholarlink

This study investigates the application of Large Language Models (LLMs) for Grammar Error Correction (GEC) and Grammar Error Explanation (GEE) in Indonesian, addressing a critical research gap for low-resource languages. We evaluate three state-of-the-art LLMs (Claude 3.7, DeepSeek R1, and LLaMA 3.2) using three reasoning techniques (Chain-of-Thought, Self-Consistency, and Tree-of-Thought) on Indonesian grammatical error correction and explanation tasks. Our comprehensive evaluation framework combines automatic metrics (precision, recall, F1, accuracy, GLEU) with human assessment by linguistic experts across multiple dimensions. Results demonstrate that Self-Consistency consistently outperforms other reasoning techniques across all models, with Claude 3.7 achieving the highest performance (F1: 0.8270). Error category analysis reveals model strengths in handling syntactic errors but persistent challenges with semantic nuances, particularly ambiguity and idiomatic expressions. We also identify a notable discrepancy between quantitative success metrics and qualitative human evaluations, suggesting new directions for assessing explanation quality in educational contexts. This research establishes a foundation for developing more effective GEC-GEE systems for Indonesian and potentially other low-resource languages. The data are available at: https://github.com/syauqiezjut/indonesian-gec-gee .

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

Exploring Large Language Models for Grammar Error Explanation and Correction in Indonesian as a Low-Resource Language

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Syauqie Muhammad Marier,
  • Daminov Bahodir,
  • Aveny Septi Astriani,
  • Linan Zhu,
  • Xiangjie Kong

摘要

This study investigates the application of Large Language Models (LLMs) for Grammar Error Correction (GEC) and Grammar Error Explanation (GEE) in Indonesian, addressing a critical research gap for low-resource languages. We evaluate three state-of-the-art LLMs (Claude 3.7, DeepSeek R1, and LLaMA 3.2) using three reasoning techniques (Chain-of-Thought, Self-Consistency, and Tree-of-Thought) on Indonesian grammatical error correction and explanation tasks. Our comprehensive evaluation framework combines automatic metrics (precision, recall, F1, accuracy, GLEU) with human assessment by linguistic experts across multiple dimensions. Results demonstrate that Self-Consistency consistently outperforms other reasoning techniques across all models, with Claude 3.7 achieving the highest performance (F1: 0.8270). Error category analysis reveals model strengths in handling syntactic errors but persistent challenges with semantic nuances, particularly ambiguity and idiomatic expressions. We also identify a notable discrepancy between quantitative success metrics and qualitative human evaluations, suggesting new directions for assessing explanation quality in educational contexts. This research establishes a foundation for developing more effective GEC-GEE systems for Indonesian and potentially other low-resource languages. The data are available at: https://github.com/syauqiezjut/indonesian-gec-gee .