摘要
针对传统文本生成摘要方法在生成摘要时存在并行能力不足以及事实性错误问题,提出在Transformer框架基础上引入事实感知的FA-TR模型。提取源文本的事实性描述信息,对该信息进行编码后与源文信息编码相融合,加入源文的位置编码一起作为编码器的输入,通过解码器对语义向量进行解码生成目标摘要。该模型改善了生成的摘要歪曲或捏造源文本事实的现象,提高了摘要质量。通过在中文短文本摘要数据集LCSTS实验,采用ROUGE作为评价指标,与其它4种方法进行实验对比与分析,验证了该模型的可行性和算法的有效性。
To solve the problems of lack of parallelism and factual errors in traditional text generation abstract methods,a FA-TR model based on Transformer framework was proposed.The factual description information of the source text was extracted,the information was encoded and it was fused with the source information encoding.The location encoding of the source text was added as the input of the encoder,and the semantic vector was decoded to generate the target abstract through the decoder.The model improves not only the distortion or fabrication of the source text,but also the quality of the abstract.Through the LCSTS experiment in the abstract,ROUGE was adopted as the evaluation index and the other four methods were compared and analyzed in experiment,which verified the feasibility of the model and the effectiveness of the algorithm.
作者
高巍
马辉
李大舟
于沛
孟智慧
GAO Wei;MA Hui;LI Da-zhou;YU Pei;MENG Zhi-hui(School of Computer Science and Technology,Shenyang University of Chemical Technology,Shenyang 110142,China;Hebei Branch of China Mobile Communication Group Design Institute Limited Company,Taiyuan 030000,China)
出处
《计算机工程与设计》
北大核心
2021年第12期3445-3452,共8页
Computer Engineering and Design
基金
辽宁省教育厅科学研究基金项目(LQ2017008)
辽宁省博士启动基金项目(201601196)
沈阳化工大学教育教学培育工程基金项目(2020,No.35)。
