SC-HyDE: Mitigating Hallucinations in Chinese Legal Question Answering via Self-Corrected Hypothetical Document Embeddings

Abstract

Retrieval-Augmented Generation (RAG) has been cited as a potential technique for reducing hallucinations in Large Language Models (LLM). However, applying RAG in the Chinese legal landscape remains a challenge due to the large semantic gap between informal user queries and professional legal terminology, and the risk of retrieval noise and false information generation. To address these concerns, this paper proposes SC-HyDE, a novel, training-free system that blends generative expansion with self-reflective filtering. First, the domain based Hypothetical Document Embeddings (HyDE) system is applied to generate an imaginative legal analysis, effectively filling out the semantic gap and restoring retrieval accuracy. Second, since HyDE can produce hallucinating legal reference points within its latent space, there is a simple module incorporated for Self-Correction Critic. This module compares the relevance of retrieved documents to the original query by using a zero-shot prompting strategy while excluding noise before the second generation stage. The proposed framework is evaluated against a sample of 588 cases from official legal datasets. The experimental results show that SC-HyDE performs well above standard RAG and vanilla HyDE levels. In particular, the method drops the Noise Rate from 88.33% to 25.06% and the Hallucination Rate from 25.68% to 13.61%, offering an effective approach for those high stakes legal consultancies where factual accuracy is a major concern.