C-LLM: Learn to Check Chinese Spelling Errors Character by Character

中文拼写检查（CSC）旨在检测和纠正句子中的拼写错误。尽管大型语言模型（LLM）表现出强大的能力并广泛应用于各种任务，但它们在CSC上的表现通常不尽如人意。我们发现，LLM未能满足CSC任务的中文字符级约束，即等长和音近，从而导致了性能瓶颈。进一步的分析揭示，这个问题源于分词的粒度，因为当前的混合字符-词分词难以满足这些字符级约束。为了解决这个问题，我们提出了C-LLM，一种基于大型语言模型的中文拼写检查方法，它学习逐个字符地检查错误。字符级分词使模型能够学习字符级对齐，有效地缓解了与字符级约束相关的问题。此外，CSC简化为复制为主、替换为辅的任务。在两个CSC基准测试上的实验表明，C-LLM相比现有方法平均提高了10%。具体而言，在一般场景下，它显示出2.1%的改进，在垂直领域场景下则显著提高了12%，建立了最先进的性能。源代码可在https://github.com/ktlKTL/C-LLM上访问。

CSC performance bottleneck caused by LLMs failing to meet Chinese character-level constraints

C-LLM, a character-level tokenization approach, enables LLMs to learn character-level alignment and effectively mitigate issues related to character-level constraints in CSC

C-LLM achieves an average improvement of 10% over existing methods, with a 2.1% improvement in general scenarios and a significant 12% improvement in vertical domain scenarios. The source code is available on GitHub at https://github.com/ktlKTL/C-LLM.

Other recent studies in this field include 'Improving Chinese Spelling Correction with Active Learning and Language Model Ensembling' and 'Chinese Spelling Correction with High-Order Word Dependency Modeling and Global Context Encoding'.