基于模型的新闻文本自动摘要研究  

Research on News Text Automatic Summarization Based on BERT Model

作　　者：陈鑫影[1] 董文超 刘月凡[1] 杨家琪 CHEN Xinying;DONG Wenchao;LIU Yuefan;YANG Jiaqi(School of Rail Intelligence Engineering,Dalian Jiaotong University,Dalian 116028,China)

机构地区：[1]大连交通大学轨道智能工程学院,辽宁大连116028

出　　处：《大连交通大学学报》2024年第5期114-120,共7页Journal of Dalian Jiaotong University

基　　金：辽宁省应用基础研究计划项目(1655706734383);辽宁省教育厅高校基本科研项目(LJKMZ20220826);辽宁省自然科学基金计划项目(2019-MS-036)。

摘　　要：针对传统文本摘要模型参数多、运算耗时长等问题,基于改进BERTSUM模型,探索了BERT模型的蒸馏变体DistilBERT、MobileBERT在CNN/DM数据集上所表现出的摘要性能,并由此提出一种轻量化抽取式摘要模型(DistilSum)。试验对比分析表明,该模型保留了原模型99%的性能,同时训练参数缩减了约36%,训练时间大幅减少。To address the issues of numerous parameters and lengthy computations in traditional text summari⁃zation models,the summarization performance on the CNN/DM dataset was explored for the distilled variants of the BERT model,such as DistilBERT and MobileBERT.Consequently,a lightweight extractive summariza⁃tion model named DistilSum is proposed.Comparative analysis indicates that the model retains 99%of the orig⁃inal model performance while reducing training parameters by approximately 36%with significantly shortening training time.

关 键 词：抽取式文本摘要 预训练语言模型 DistilBERT BERTSUM 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]