摘要
大型语言模型在自然语言处理任务中取得显著进展,但其对封装在参数内的知识依赖易引发幻觉现象。为缓解这一问题,检索增强生成技术通过信息检索方法降低错误风险。然而,现有方法检索到的文档往往含有不准确或误导性信息,且在评估文档相关性方面存在判别准确性不足的问题。针对上述挑战,设计了一种简洁高效的方法,通过结合稀疏检索与稠密检索,兼顾词汇重叠的信息与语义相关性。此外,引入排序器对检索到的候选段落进行重排序,在排序器的输入中注入稀疏和稠密检索的分数,进一步优化了段落的排序质量。为验证所提方法的有效性,在SQuAD和HotpotQA数据集上进行实验,并与现有基准方法比较。实验结果表明,所提方法在提升问答性能方面具有显著优势。
Large language models have made significant progress in natural language processing tasks,but their reliance on knowledge encapsulated within parameters can easily lead to the phenomenon of hallucinations.To mitigate this issue,retrieval-augmented generation techniques reduce the risk of errors through information retrieval methods.However,existing methods often retrieve documents that contain inaccurate or misleading information,and there is a lack of discriminative accuracy in evaluating document relevance.In response to these challenges,this study designs a concise and efficient method that combines sparse retrieval with dense retrieval,taking into account both lexical overlap and semantic relevance.Furthermore,a ranker is introduced to reorder the retrieved candidate paragraphs,with the input to the ranker infused with scores from both sparse and dense retrieval,further optimizing the quality of paragraph ranking.To validate the effectiveness of this method,experiments were conducted on the SQuAD and HotpotQA datasets,and comparisons were made with existing benchmark methods.The experimental results demonstrate that this method holds a significant advantage in enhancing question-answering performance.
作者
白云天
郝文宁
靳大尉
BAI Yuntian;HAO Wenning;JIN Dawei(College of Command&ControlEngineering,Army Engineering University of PLA,Nanjing 210000,China)
出处
《计算机科学》
北大核心
2025年第S1期36-42,共7页
Computer Science
基金
国防工业技术发展计划(JCKY2020601B018)。
关键词
大型语言模型
检索增强生成
信息检索
Large language model
Retrieval-augmented generation
Information retrieval
