【作者】 郭建强；

【导师】 任鹏杰；

【作者基本信息】 山东大学 ， 计算机技术（专业学位）， 2025， 硕士

【摘要】 随着信息技术的快速发展,新闻行业既迎来机遇,也面临挑战。传统新闻撰写效率低,新闻搜索方式也难以满足用户的深入分析需求。大语言模型的出现,为新闻内容生产及用户获取新闻信息提供了新思路,不仅能够提升新闻撰写效率,还能帮助用户高效获取新闻核心内容。然而,大语言模型在新闻领域直接应用存在信息滞后导致的错误信息传播问题。由于其训练数据主要基于历史信息,模型难以实时同步最新的新闻动态,这在很大程度上影响了新闻报道的时效性,进而可能导致错误信息的传播。为解决上述问题,可以采用检索增强技术和模型编辑技术。检索增强技术可以从外部知识库中获取最新的信息并提供给模型,而模型编辑技术则可以直接对模型内部的信息进行更新。但是,这些现有方法也面临着各自的挑战。在检索增强方面,现有方法主要面临难以深入理解用户查询意图、难以同时提升检索的精确率和召回率两个挑战。在模型编辑方面,现有方法主要面临难以提高模型对编辑内容的泛化性、难以避免编辑次数增多引发的知识遗忘或效率下降问题、难以实现对新知识的快速编辑三个挑战。同时,如果仅依赖检索增强方法从外部补充知识,虽然模型能够获取新的信息,但难以泛化理解新知识,只能被动地应用这些知识。而如果仅通过模型编辑方法更新内部知识,又会受限于模型的存储容量,无法更新大量的新知识。因此,本文从模型的外部增强和内部优化两个角度出发,设计了两个核心算法:基于多源融合的检索增强生成算法(MF-RAG)和基于自适应检索与引导学习的模型编辑算法(SRGL-ME),以更好地应对现有方法面临的挑战。具体而言,MF-RAG算法通过两个关键步骤应对用户查询意图难以理解的挑战:一是重构问题,结合对话历史清晰捕捉用户意图,生成准确查询语句;二是生成假设性文档,通过深度分析问题拓宽检索语义空间,提高获取高质量文档的概率。针对检索精确率和召回率难以同时提升的问题,MF-RAG算法设计双阶段检索策略:第一阶段采用多路召回策略扩大候选文档规模,提高召回率;第二阶段通过重排序技术筛选高相关性文档,提升准确率。SRGL-ME算法针对模型编辑时知识遗忘或效率下降的问题,采用知识缓存法避免直接修改模型内部参数,同时防止引入额外参数导致效率下降。为提高模型对编辑内容的泛化理解,设计知识表示模块,将新知识转化为深层次语义知识存储。针对新知识快速编辑问题,设计自适应检索模块,通过可训练的自适应阈值参数动态调整阈值,自动适应新知识变化。此外,通过校正提示生成模块和知识融合模块解决知识前缀过长影响模型主体任务理解的问题,并设计损失函数引导模型纠正内部错误知识。为了充分发挥本文所设计的MF-RAG和SRGL-ME两种算法的优势和实际应用价值,本文基于这两种算法构建了一个智能新闻问答系统,主要面向新闻从业人员和普通大众用户。该系统不仅为用户提供便捷的新闻问答服务,还能通过管理员操作实时更新知识库和模型知识,提升系统的实用性和准确性。通过算法优化和系统实现,本文旨在提高大语言模型生成新闻内容的质量和数据安全性,同时支持模型知识的快速更新,为新闻媒体行业提供更高效的信息传播渠道和更广泛的应用场景。更多还原

【Abstract】 With the rapid development of information technology,the news industry faces both opportunities and challenges.Traditional news writing is often inefficient,while conventional news search methods fail to meet users’ needs for in-depth analysis.The emergence of large language models(LLMs)provides new possibilities for news content production and information retrieval,significantly improving writing efficiency and helping users quickly grasp key news insights.However,the direct application of large language models(LLMs)in the news domain faces the issue of misinformation propagation due to information lag.Since their training data primarily relies on historical information,these models struggle to synchronize with real-time news updates,significantly compromising the timeliness of news reporting and potentially leading to the spread of misinformation.To address these challenges,retrieval-augmented techniques and model editing methods can be employed.Retrieval-augmented techniques enable the model to access up-to-date information from external knowledge bases,while model editing allows direct updates to the model’s internal knowledge.Nevertheless,existing approaches encounter their own limitations.For retrieval-augmented methods,key challenges include difficulties in deeply understanding user query intent and simultaneously improving both retrieval precision and recall.For model editing,major obstacles involve enhancing the generalization of edited knowledge,mitigating knowledge forgetting or efficiency degradation caused by frequent edits,and enabling rapid updates for new knowledge.Moreover,relying solely on retrieval-augmented methods to supplement external knowledge may allow the model to acquire new information but fails to generalize its understanding,limiting the model to passive utilization of such knowledge.Conversely,depending exclusively on model editing to update internal knowledge is constrained by the model’s storage capacity,preventing the integration of large-scale new knowledge.To overcome these limitations,this paper proposes two core algorithms from the perspectives of external enhancement and internal optimization:the Retrieval-Augmented Generation Algorithm Based on Multi-source Fusion(MF-RAG)and the Model Editing Algorithm Based on Self-adaptive Retrieval and Guided Learning(SRGL-ME).These algorithms aim to address the existing challenges in current methods.Specifically,MF-RAG tackles the difficulty of understanding user query intent through two key steps:(1)Query Reconstruction,which captures user intent by integrating dialogue history to generate precise queries;(2)Hypothetical Document Generation,which expands the retrieval semantic space through deep analysis to increase the probability of obtaining high-quality documents.To simultaneously improve both retrieval precision and recall,MF-RAG employs a two-stage retrieval strategy:the first stage uses a multi-route recall strategy to expand the candidate document pool and enhance recall,while the second stage applies re-ranking techniques to filter highly relevant documents,improving precision.The SRGL-ME algorithm addresses knowledge forgetting and efficiency decline during model editing by adopting a knowledge caching strategy,preventing direct modifications to model parameters while avoiding efficiency losses caused by additional parameters.To enhance the model’s generalization of edited knowledge,a knowledge representation module is designed to convert new knowledge into deep semantic representations for storage.For rapid new knowledge updates,an adaptive retrieval module is introduced,which dynamically adjusts thresholds using trainable adaptive parameters to accommodate changes in new knowledge.Additionally,a correction-based prompt generation module and a knowledge fusion module are designed to mitigate the impact of long knowledge prefixes on the model’s primary task understanding.A loss function is also formulated to guide the model in correcting internal erroneous knowledge.To fully utilize the advantages of the proposed MF-RAG and SRGL-ME algorithms,this paper develops an intelligent news question-answering system designed for both journalists and the general public.This system not only provides users with efficient news Q&A services but also allows administrators to update the knowledge base and model knowledge in real time,enhancing its practicality and accuracy.Through algorithmic optimization and system implementation,this study aims to improve the quality and security of LLM-generated news content while enabling rapid model knowledge updates.Ultimately,this work offers a more efficient information dissemination channel and broader application scenarios for the news media industry.更多还原