基于文本语义的注意力指针网络文本摘要生成模型  

Text Semantic-based Multi-attention Pointer Network Text Summary Generation Model

作　　者：谢文博 张晓滨[1] XIE Wenbo;ZHANG Xiaobin(School of Computer Science,Xi'an Polytechnic University,Xi'an 710048)

机构地区：[1]西安工程大学计算机科学学院,西安710048

出　　处：《计算机与数字工程》2025年第1期189-195,共7页Computer & Digital Engineering

摘　　要：论文旨在针对文本摘要生成任务中存在的语义信息编码不充分、生成摘要结果不通顺问题,提出一种基于文本语义的注意力指针网络文本摘要模型。该模型采用改进的序列到序列(Seq2Seq)架构,利用双编码器+双注意力机制对源文档编码以获取文本的不同特征向量:应用Child-Sum Tree-LSTMs+SelfAttention获取文本的语义特征向量,BiLSTM+SoftAttention获取文本的位置时序特征向量,之后构建门控机制与指针网络融合取舍不同编码器获取到的特征向量,利用覆盖机制解决生成重复问题,最后使用集束搜索选取最终生成词,从而产生更为准确和连贯的摘要。最终实验表明:在中文短文本摘要数据集LCSTS与英文数据集CNN/Daily Mail上,论文模型与对照实验组对比,在ROUGE评分标准下取得了更高的分数,表明该模型能有效地提升文本摘要生成效果。This paper aims to solve the problems of insufficient semantic coding and unsmooth summary sentences in the task of text summary generation,and proposes an attention pointer network text summary model based on text semantics.The model adopts the improved sequence-to-sequence(Seq2Seq)architecture,and uses double encoders and double attention mechanism to encode the source document to obtain different feature vectors of the text.The semantic feature vectors of the text are obtained by us⁃ing Child-Sum Tree-LSTMs+Self Attention,and the position and time sequence feature vectors of the text are obtained by using BiLSTM+SoftAttention.After that,the gating mechanism is constructed and the pointer network is fused to select the feature vectors obtained by different encoders,and the covering mechanism is used to solve the problem of generating repetition.Finally,the clus⁃ter search is used to select the final generated words,so as to generate a more accurate and coherent summary.Final experiment shows that,compared with the control experimental group,the model in this paper gets a higher score under the ROUGE scoring standard on the Chinese short text abstract data set LCSTS and the English data set CNN/Daily Mail,which proves that the model can effectively improve the quality of text abstracts.

关 键 词：文本摘要生成 Child-Sum Tree-LSTMs Seq2Seq 指针网络 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]