Knowledge-Augmented Question Error Correction for Chinese Question Answer System with QuestionRAG

Longpeng Qiu, Ting Li, Shuai Mao, Nan Yang, Xiaohui YanEMNLP 2025

TL;DR

This paper introduces QuestionRAG, a novel framework that significantly improves Chinese question error correction for QA systems. By combining Retrieval-Augmented Generation (RAG) with Reinforcement Learning (RL), QuestionRAG effectively addresses both misinterpretation and over-correction issues in Large Language Models (LLMs), leading to more accurate and reliable question understanding.

What Problem Does This Paper Address?

Question Answering (QA) systems are fundamentally limited by errors in user questions. These errors, stemming from sources like Automatic Speech Recognition (ASR), Optical Character Recognition (OCR), or user typos, can lead to incorrect responses. While Large Language Models (LLMs) show promise for correcting these errors, they struggle with two main issues:

  1. Misinterpretation: LLMs may fail to understand the user's intent due to a lack of specific background knowledge, especially for nuanced terms or proper nouns.
  2. Over-correction: LLMs can be too aggressive, correcting grammatically valid but less common phrases into more conventional ones, thereby altering the original meaning or intent.

Existing methods often rely on task-specific models or handcrafted rules, limiting their generalizability. This paper aims to develop a unified framework that can handle a broad spectrum of errors without extensive customization.

Methods

The proposed framework, QuestionRAG, tackles these problems with a two-pronged strategy:

  1. Knowledge Augmentation (for Misinterpretation): QuestionRAG employs Retrieval-Augmented Generation (RAG) to enrich the input question with external knowledge. This includes:
  2. Web Search Results: Titles of relevant web pages.
  3. QA System Logs: Similar questions previously encountered.
  4. Wikipedia: Entities and their descriptions. This retrieval process considers multiple similarity factors (lexical, semantic, phonetic) to gather relevant context, helping the LLM understand nuanced queries.

  5. Reinforcement Learning (RL) for Alignment (for Over-correction): To prevent over-correction and ensure the LLM adheres to precise correction, QuestionRAG uses a reinforcement learning training method. Specifically, it employs Group Relative Policy Optimization (GRPO). This approach trains the LLM to generate corrections by comparing candidate responses and optimizing a reward function that balances:

  6. Format Rewards: Ensures the output follows a specific structure (e.g., <think>{reasoning}</think>\n\n{final_answer}).
  7. Accuracy Rewards: Measures the similarity between the generated correction and the ground truth using a normalized edit distance, penalizing regressions.

The overall training process involves a cold-start Supervised Fine-Tuning (SFT) phase followed by GRPO-based reinforcement learning for post-training.

Datasets

The paper utilizes the following datasets for training and evaluation:

  • QCSet: A proprietary dataset of approximately 10,000 error-correction pairs, manually annotated for high accuracy.
  • MCSCSet: An open-source dataset for medical Chinese spelling correction, containing about 200,000 samples.
  • Qspell: A public-domain dataset of approximately 250,000 error-correction pairs, covering various topics and languages.

The training data was augmented with reasoning traces generated by DeepSeek R1.

Key Insights and Ideas

  • Knowledge is King (Even for LLMs): The paper strongly demonstrates that augmenting LLMs with external knowledge (via RAG) is crucial for overcoming misinterpretation, even more so than simply increasing model size. A smaller LLM with QuestionRAG can outperform a larger LLM without it.
  • RL for Precise Control: Reinforcement learning, particularly GRPO, is highly effective in aligning LLMs for the specific task of question correction, significantly mitigating over-correction and improving generalization.
  • Unified Framework for Diverse Errors: QuestionRAG offers a unified approach to handle various error types, moving beyond specialized models for phonetic or grammatical errors.
  • Robustness to Noisy Context: The GRPO-trained QuestionRAG model shows remarkable robustness, successfully correcting ambiguous queries even when the retrieved contextual information is noisy or misleading.

Key Results

  • Quantitative Improvements: QuestionRAGGRPO achieved the lowest Character Error Rate (CER) across all tested datasets (QCSet, MCSCSet, Qspell), with CERs of 9.04%, 5.08%, and 6.37% respectively. This significantly outperforms baselines like "No RAG" (Qwen3-8B) and even fine-tuned T5 models.
  • SFT vs. RL: Both SFT and RL improved performance over the RAG-only approach (QuestionRAGICL). QuestionRAGSFT slightly outperformed QuestionRAGICL, while QuestionRAGGRPO achieved the best results, indicating the superiority of RL for generalization.
  • Impact of Knowledge Augmentation: Table 3 shows a consistent improvement in CER as more types of knowledge (entities, web info, similar questions) are added to the RAG component, even for the "No RAG" baseline.

Limitations and Future Work

  • Dependency on Search Quality: The performance of QuestionRAG is critically dependent on the quality and relevance of the retrieved search results. Noisy or irrelevant information can compromise the final output.
  • Increased Latency: The RAG process, by increasing prompt length, can lead to higher latency, particularly for the time to first token.
  • Future Work: The authors suggest that the QuestionRAG methodology can be extended to broader NLP tasks such as question rewriting, query planning, and enhancing query understanding in search engines. Further research could explore more sophisticated retrieval and ranking mechanisms.

Summary updated

BibTeX
@inproceedings{qiu2025knowledge,
  author = {Longpeng Qiu and Ting Li and Shuai Mao and Nan Yang and Xiaohui Yan},
  title = {{Knowledge-Augmented Question Error Correction for Chinese Question Answer System with QuestionRAG}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2025},
}
Figure