Scholarlink

<p>Chinese grammatical error correction remains challenging due to flexible word order, implicit morphology, and the need for context-dependent reasoning. Most methods still treat correction as a direct sequence transformation and learn to mimic edits without understanding why they are correct, which leads to spurious modifications and unstable performance across diverse grammatical patterns. This paper introduces RICE, a reasoning-integrated framework that models correction as a reasoning-conditioned decision process. RICE employs multi-agent reflective distillation to acquire structured grammatical cognition, reasoning-mediated decoder integration to inject explanatory signals into edit generation, and self-evolving reinforcement to refine edit reliability under imperfect supervision. Experiments on two public datasets show that RICE achieves competitive precision-oriented correction performance. These results demonstrate the benefit of aligning reasoning trajectories with correction behavior and support a shift toward reasoning-centered grammatical error correction.</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

RICE: Reasoning-integrated correction evolution for Chinese grammatical error correction

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月24日

查看全文
  • Yue Yang,
  • Yongbin Qin,
  • Ruizhang Huang,
  • Ruina Bai

摘要

Chinese grammatical error correction remains challenging due to flexible word order, implicit morphology, and the need for context-dependent reasoning. Most methods still treat correction as a direct sequence transformation and learn to mimic edits without understanding why they are correct, which leads to spurious modifications and unstable performance across diverse grammatical patterns. This paper introduces RICE, a reasoning-integrated framework that models correction as a reasoning-conditioned decision process. RICE employs multi-agent reflective distillation to acquire structured grammatical cognition, reasoning-mediated decoder integration to inject explanatory signals into edit generation, and self-evolving reinforcement to refine edit reliability under imperfect supervision. Experiments on two public datasets show that RICE achieves competitive precision-oriented correction performance. These results demonstrate the benefit of aligning reasoning trajectories with correction behavior and support a shift toward reasoning-centered grammatical error correction.