摘要
在实际业务中时,常面临文本与它对应的其他模态在时间响应上难以同步的问题。例如,数字人实时手语表演无法与新闻口播同步播放。为了解决长度可控问题,提出一种基于大语言模型(LLM)的文本摘要解决方案,旨在保持原文语义不变的前提下将文本压缩至指定长度。首先通过模板调优和人工评估的方式,确定最适合长度可控文本摘要的LLM和模板;在此基础上,利用ChatGPT得到一定量优质的满足长度需求的文本摘要训练样本;其次,结合低秩自适应微调(LoRA)技术,利用生成的数据样本集对选定的大语言模型Baichuan-13B-Chat进行微调。在推理阶段,通过微调后的LLM生成多个结果和文本筛选模块打分,最终得到语义相对完整且长度满足要求的摘要文本。实验结果表明,所提方案在亚运手语新闻数据中指标显著提升,人工评估的平均满意度达到88.53%,整体压缩达标率达到73.7%,基本满足实际生产应用的需求。
In practical business scenarios,it is often challenging to synchronize text with other modalities in terms of temporal response.For instance,realtime sign language performance by digital avatars may not synchronize with the speed of news broadcasts.To address the above-mentioned issue,a solution was proposed for text summarization system based on Large Language Model(LLM)to compress the text to a specified length while preserving the original semantic meaning.First,the most suitable LLM and prompt for text summarization were determined by prompt tuning and manual evaluation.On this basis,ChatGPT was employed in generating a substantial amount of high-quality dataset that meet the length requirements.Then,Low-Rank Adaptation(LoRA)technology was used in fine-tuning the selected LLM,Baichuan-13B-Chat,with the generated dataset.During the inference stage,multiple results were generated by the fine-tuned LLM and scored by text filtering module to ultimately obtain semantically relatively complete and length-compliant summary.Experimental results showed that the proposed approach significantly improved metrics for the Asian Games sign language news.The average satisfaction rate from manual evaluations reached 88.53%,and the compression compliance rate reached 73.7%,effectively meeting the practical requirements for production applications.
作者
黄君豪
朱锦文
向宗元
李萌坚
毛瑞琛
HUANG Junhao;ZHU Jinwen;XIANG Zongyuan;LI Mengjian;MAO Ruichen(Zhejiang Lab,Hangzhou Zhejiang 311121,China)
出处
《计算机应用》
CSCD
北大核心
2024年第S01期29-33,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
文本摘要
长度可控
大语言模型
低秩自适应微调
模板调优
文本筛选
text summarization
length controllability
Large Language Model(LLM)
Low-Rank Adaptation(LoRA)
prompt tuning
text filtering
