CiteCheck: Towards Accurate Citation Faithfulness Detection
CiteCheck introduces a large-scale Chinese citation faithfulness dataset using a two-stage manual annotation and LLM-based data augmentation, enabling efficient training of models with high robustness.
Key Findings
Methodology
The study constructs CiteCheck with over 11,000 samples, employing a two-stage annotation process: first, professional annotators label support relations; second, LLM-generated negative samples are manually filtered for quality. Data augmentation uses GPT-4o to modify supporting documents, creating diverse negative samples at low cost (86% reduction). Fine-tuning models with LoRA on 7B-8B architectures yields high performance. Experiments reveal the original samples are challenging, with even state-of-the-art LLMs achieving limited accuracy (~70%).
Key Results
- On the original test set, GPT-4o achieves 83.7% overall accuracy, but only 70.4% on negative samples, indicating high difficulty. Fine-tuned smaller models like Llama-3.1-8B reach over 90% accuracy, demonstrating the effectiveness of augmented data. Negative samples generated via information modification and deletion ensure diversity and realism.
- Models trained on augmented data outperform baseline models, especially in negative sample detection, with Llama-3.1-8B achieving 90.6% accuracy, validating the data augmentation strategy's utility.
- Analysis shows that models trained with our dataset generalize well to challenging real-world samples, significantly improving robustness in citation support detection tasks.
Significance
This work addresses the scarcity of Chinese citation faithfulness datasets, providing a large, high-quality benchmark. The innovative data augmentation reduces annotation costs substantially, facilitating scalable model training. It advances research in Chinese NLP, supports verification in academic and knowledge graph applications, and promotes trustworthy AI systems. The dataset and methods lay a foundation for future exploration of multi-label and multi-modal citation verification, impacting both academia and industry.
Technical Contribution
The main technical innovation is the two-stage annotation combined with LLM-based negative sample generation, drastically reducing costs while maintaining quality. The use of GPT-4o for document modification simulates real-world errors, and LoRA fine-tuning enhances efficiency on smaller models. This integrated approach offers a scalable, effective solution for Chinese citation verification, with potential for adaptation to other languages and tasks.
Novelty
This is the first large-scale Chinese dataset for citation faithfulness detection employing a two-stage annotation process with LLM-generated negative samples. Unlike prior English datasets, CiteCheck emphasizes cost-effective, diverse negative sample creation through document modification, and demonstrates the effectiveness of parameter-efficient fine-tuning in this context. It bridges resource gaps and introduces a novel methodology for low-resource environments.
Limitations
- The dataset only provides binary labels, lacking fine-grained evidence annotations, which limits interpretability and detailed analysis.
- Experiments are mainly confined to specific Chinese datasets, so generalization across domains remains to be validated.
- Reliance on LLM for negative sample generation may introduce biases or inconsistencies, affecting diversity and realism.
Future Work
Future directions include expanding to multi-label and evidence-level annotations, integrating multi-modal data (images, structure), and developing explainable models. Additionally, exploring more sophisticated negative sample generation strategies and cross-lingual transfer learning could further improve robustness and applicability.
AI Executive Summary
Ensuring the trustworthiness of automatically generated content is vital in NLP, especially for academic and knowledge-based applications. While English datasets for citation verification have matured, Chinese resources remain scarce, hindering progress. This paper introduces CiteCheck, a large-scale Chinese dataset with over 11,000 samples, designed to evaluate citation faithfulness. The dataset’s construction leverages a two-stage manual annotation process: first, professional annotators label support relations; second, LLM-generated negative samples are filtered for quality. To address the high cost of negative sample annotation, the authors employ GPT-4o to modify documents, creating diverse, realistic negative samples at a fraction of traditional costs (86% savings). The training data, combined with parameter-efficient fine-tuning (LoRA), enables smaller models (7B-8B) to achieve over 90% accuracy, outperforming zero-shot large models like GPT-4o, which achieve around 70%. The original samples pose significant challenges, emphasizing the need for robust datasets. The results demonstrate that augmented data significantly enhances model robustness, especially in negative sample detection, crucial for real-world applications such as academic verification and knowledge graph construction. This work not only provides a valuable resource for Chinese NLP but also offers a scalable, low-cost methodology for dataset construction, with promising avenues for future multi-label and multi-modal extensions.
Deep Analysis
Background
随着大规模语言模型（LLMs）在自然语言处理中的广泛应用，确保内容真实性成为关键挑战。检索增强生成（RAG）系统通过结合外部知识源，提升回答的可靠性，但引文的真实性仍存在问题。英文引文检测数据集逐渐丰富，但中文资源稀缺，限制了研究发展。构建高质量数据集面临标注成本高、负样本稀缺等难题，促使研究者探索低成本、高效率的解决方案。近年来，LLM辅助数据增强逐渐成为热点，为低资源环境下的引文验证提供新思路。
Core Problem
核心问题在于缺乏大规模、标注高质量的中文引文可信性数据集，尤其是负样本的稀缺严重制约模型性能。传统标注方式耗时耗力，难以获得真实错误样本，限制模型在实际场景中的表现。负样本的多样性不足也影响模型泛化能力，导致在复杂应用中识别能力不足。这些问题阻碍了中文引文验证技术的推广和应用。
Innovation
本研究的创新点包括：1）提出两阶段人工标注策略，结合LLM生成负样本，显著降低成本（86%节省）；2）利用GPT-4o对支持信息进行微调，生成多样化负样本，确保数据质量；3）采用参数高效微调（LoRA），在较小模型上实现优异性能，突破了传统高成本训练的限制。这一方案结合了人机协作和技术创新，为低资源环境下的引文检测提供了新路径。
Methodology
- �� 收集多源中文问答数据，模拟真实RAG场景，生成带引用的回答。
- �� 由专业标注员判断样本支持关系，形成正负样本基础。
- �� 利用GPT-4o对支持信息进行微调，删除或修改关键段落，生成负样本。
- �� 两阶段筛选：第一阶段由标注员判断原始样本正负关系，第二阶段筛选生成的负样本质量。
- �� 最终，结合正负样本，采用LoRA微调模型，提升检测性能。
Experiments
在CiteCheck的测试集上，评估多种零-shot大型模型（如GPT-4o、Qwen2.5-Plus）表现，发现其在负样本识别上表现有限（准确率约70%）。通过微调较小模型（如Llama-3.1-8B），利用增强数据，准确率提升至90%以上。实验还分析了不同负样本生成策略对模型性能的影响，验证了信息修改的有效性。模型在实际场景中展现出较强鲁棒性，验证了数据增强策略的实用性。
Results
原始测试样本难度大，GPT-4o准确率为83.7%，负样本识别为70.4%。增强数据训练后，Llama-3.1-8B模型在测试集上达90.6%的整体准确率，负样本识别率也显著提升。负样本由信息修改和删除两种方法生成，确保多样性。模型在识别不支持引文方面表现优异，有效提升了系统的可靠性。
Applications
该数据集和方法适用于学术搜索、知识图谱构建、自动引文验证等场景。模型可在科研、学术出版、智能问答等领域部署，提升内容真实性和可信度。依赖低成本数据增强策略，适合资源有限的应用环境，推动行业标准化发展。
Limitations & Outlook
目前仅提供二元标签，未标注具体支持证据，限制模型解释能力。实验主要在特定中文数据集上，泛化性待验证。负样本生成依赖LLM，可能存在偏差，影响多样性。未来需引入多标签和多模态信息，提升模型细粒度理解和适应性。
Plain Language Accessible to non-experts
想象你在厨房里做菜，食材代表信息，调料代表证据。做菜时，你需要确认每个调料是否真正让菜更好吃。如果调料不合适，菜就会变得难吃。引文检测就像是判断每个引用是否真正支持文章的内容。传统方法像是用手一一检查，费时又繁琐。现在，有了智能厨师（LLM），它可以帮你快速修改调料，生成不同版本的菜谱（负样本），让你训练厨艺（模型）变得更快更好。最终，你的厨艺（模型）能更准确地判断哪些调料（引用）是真正支持菜（内容）的，哪些是虚假的。
ELI14 Explained like you're 14
想象你在学校的厨房里做饭，你用各种调料（信息）来让菜变得更好吃。有时候，你会用错调料，比如盐放多了，菜就会变得难吃。引文检测就像是检查每个引用是不是在说菜的真话。以前，要一个个检查很麻烦，还容易出错。现在，有个聪明的机器人（LLM），它可以帮你试验不同的调料组合，告诉你哪些是真的能让菜变得更好，哪些是虚假的。通过让机器人帮忙修改调料（负样本），你可以训练你的厨艺（模型）变得更厉害，能更快分辨出哪些引用是真的支持内容，哪些是虚假的。这样，你的菜（内容）就会变得更可靠，大家也更相信你的厨艺了。
Abstract
Citation faithfulness detection is critical for enhancing retrieval-augmented generation (RAG) systems, yet large-scale Chinese datasets for this task are scarce. Existing methods face prohibitive costs due to the need for manually annotated negative samples. To address this, we introduce the first large-scale Chinese dataset CiteCheck for citation faithfulness detection, constructed via a cost-effective approach using two-stage manual annotation. This method balances positive and negative samples while significantly reducing annotation expenses. CiteCheck comprises training and test splits. Experiments demonstrate that: (1) the test samples are highly challenging, with even state-of-the-art LLMs failing to achieve high accuracy; and (2) training data augmented with LLM-generated negative samples enables smaller models to attain strong performance using parameter-efficient fine-tuning. CiteCheck provides a robust foundation for advancing citation faithfulness detection in Chinese RAG systems. The dataset is publicly available to facilitate research.