文献类型：期刊文献

中文题名：基于深度学习的子句级文本摘要模型

英文题名：Clause-level Text Summarization Model Based on Deep Learning

作者：房一泉[1,2];沈斌[1,2];程华[2];杜嘻嘻[2]

机构：[1]华东理工大学信息化办公室,上海200237;[2]华东理工大学信息科学与工程学院,上海200237

年份：2024

卷号：52

期号：7

起止页码：2135

中文期刊名：计算机与数字工程

外文期刊名：Computer & Digital Engineering

收录：CSTPCD

基金：赛尔网络下一代互联网技术创新项目(编号:NGII20170520)资助。

语种：中文

中文关键词：文本摘要;深度学习;子句级;依存句法;ROUGE评测

外文关键词：text summarization;deep learning;clause-level;dependency parsing;ROUGE evaluation

摘要：针对文本自动摘要任务中整句级抽取式模型存在摘要过于冗余,以及训练目标与评价目标不匹配的问题,论文提出了一种基于深度学习的子句级文本摘要模型(CS-ASum)。首先,基于依存句法从原文中抽取子句级单元;然后,利用基于BERT预训练模型和基于Transformer模型的编码器获得子句的向量表示,得到初步候选摘要;最后,通过摘要匹配器计算候选摘要和原文的语义得分,得到最佳摘要。在CNN/Daily Mail数据集上的实验结果表明,CS-ASum在自动评测和人工评测中均优于对比模型,相较于表现最好的生成式摘要模型和抽取式摘要模型,CS-ASum的平均ROUGE指标值分别提高了0.76%和1.07%,由此可见,CS-ASum模型在自动文本摘要任务中比基础模型获得了更简洁、更忠于原文的摘要。
The sentence-level extractive summarization model is relatively redundant and the training objectives do not match the evaluation goals in text automatic summarization task.In order to solve these problems,a clause-level text summarization model based on deep learning(CS-ASum)is proposed.Firstly,the clause-level units of the original text are extracted based on dependen?cy parsing method.Then,the vector representation of the clause are obtained by using the encoders based on BERT pre-trained model and Transformer model,and the preliminary summary candidates are received.Finally,the summary matcher is used to cal?culate the semantic similarity between the summary candidates and the original text,and the best summary is obtained.The experi?mental results on CNN/Daily Mail dataset show that,CS-ASum model is better than the compared models in both automatic evalua?tion and manual evaluation.Compared with the best-performing abstractive summarization model and extractive summarization mode,the average ROUGE value of CS-ASum model is increased by 0.76%and 1.07%respectively.Therefore,CS-ASum model acquires more concise and comprehensive summaries than the basic models in automatic text summarization task.