DISCERN: Chain-of-Thought-Augmented Syntactic-based In-Context Learning for Chinese Semantic Error Detection
Bowen Ruan, Hongyan Wu, Yitong Han, Shengyi Jiang, Lianxi Wang, Nankai Lin · Data Intelligence · 2025
Semantic errors in Chinese text can significantly impact text comprehension and information accuracy. Although semantic error detection is crucial for improving the quality and reliability of texts, existing models still face major challenges in detection performance due to the diversity and complexity of semantic error types. To address this issue, we propose a chain-of-thoughtaugmenteD syntactIc-baSed in-ContExt LeaRNing framework (DISCERN), which aims to enhance the performance of Large Language Models’ (LLMs) in Chinese semantic error detection tasks. DISCERN consists of three core modules: Semantic Error Mechanism Mining (SEMM), Demonstration Sample Selection (DSS), and Template Filling Output (TFO). It integrates Chainof- Thought (CoT) prompting technique of LLMs with dependency syntax tree-based similarity calculation to select appropriate demonstration examples. Experiments on the CSED-R dataset demonstrate that compared to existing methods, DISCERN can effectively improve the performance of LLMs in semantic error detection tasks.