摘要
现有大语言模型(LLM)由于存在信息滞后性,在特定风格新闻稿件生成任务上存在生成内容捏造、行文不流畅连贯等问题。为了缓解这些问题,提出一套基于实时在线的web_search技术和量化低秩适配器(QLoRA)微调技术的新闻文稿生成系统的解决方案。首先,利用Bing和Google提供的API根据给定的新闻标题,获取最新的新闻素材集合;其次,利用语义相关性模型和摘要模型对初始素材集合进行筛选和文本处理,选取准确的新闻内容;再次,设计动态的prompt模板综合处理检索到的新闻素材,并在prompt中加入新闻风格约束提示词;最后,将完整的prompt提示词指令输入经过QLoRA微调的LLM中,生成新闻文稿。实验结果显示,在人工整理的热点新闻标题数据集上,所提方案生成的新闻在内容正确性、逻辑连贯性等多维人工评估标准上的平均准确率达到90%,满足实际生产应用的需求,有效提高了新闻生产的效率和质量。目前,该系统已在杭州文广集团内部成功部署应用。
The existing Large Language Models(LLMs)suffer from fabrication of generated content and insufficient flow coherence in the task of generating a specific style of expression release due to the information lag.To address these problems,a news manuscript generation system was proposed based on real-time information retrieval technology and Quantized Low-Rank Adapter(QLoRA)fine-tuning technology.Firstly,the latest news were collected based on the given news title by accessing Bing and Google API.Subsequently,a semantic relevance model and an abstract model were employed to filter the initial material set,selecting highly relevant and accurate contents.Secondly,a dynamic prompt template was designed to process the retrieved news materials,incorporating the corresponding news style as the constraint prompts.Finally,the complete prompt instructions were input into the LLM that had been fine-tuned with QLoRA to generate the manuscripts.The experimental results demonstrate that the proposed solution achieves an average accuracy of 90%in generating news manuscripts according to a manually curated dataset of hot news headlines.The validation is based on multidimensional human evaluation criteria,including content correctness,logical coherence and so on.The proposed system design can meet the requirements of practical production applications,and improve the efficiency and quality of news production.Currently,the system has been successfully deployed in Hangzhou Culture,Radio and Television Group.
作者
励琦
刘志强
李岚
向宗元
毛瑞琛
陈群
LI Qi;LIU Zhiqiang;LI Lan;XIANG Zongyuan;MAO Ruichen;CHEN Qun(Hangzhou Culture,Radio and Television Group,Hangzhou Zhejiang 311121,China;Zhejiang Lab,Hangzhou Zhejiang 311121,China)
出处
《计算机应用》
CSCD
北大核心
2024年第S01期34-38,共5页
journal of Computer Applications
基金
之江实验室跨媒体智能短视频生成关键技术项目(108001-AC2101)。
关键词
在线检索
量化低秩适配器
微调范式
大语言模型
文稿生成
提示词
online retrieval
Quantized Low-Rank Adapter
fine tuning paradigm
Large Language Model(LLM)
manuscript generation
prompt
