基于距离注意力的双特征融合的图书分类方法  

The Dual-Feature Fusion Book Classification Method Based on Distance-Attention Mechanism

作　　者：王豹 周勇[2] 褚庆福 罗淇丰 吴佳隆 许婕 WANG Bao;ZHOU Yong;CHU Qingfu;LUO Qifeng;WU Jialong;XU Jie(College of Digital Industry,Jiangxi Normal University,Shangrao Jiangxi 334000,China;School of Computer and Information Engineering,Jiangxi Normal University,Nanchang Jiangxi 330022,China;School of Software,Jiangxi Normal University,Nanchang Jiangxi 330027,China)

机构地区：[1]江西师范大学数字产业学院,江西上饶334000 [2]江西师范大学计算机信息工程学院,江西南昌330022 [3]江西师范大学软件学院,江西南昌330027

出　　处：《江西师范大学学报(自然科学版)》2024年第6期652-660,共9页Journal of Jiangxi Normal University(Natural Science Edition)

基　　金：江西省教育厅重点教学改革课题(JXJG-22-2-15);江西省教育厅科学技术研究基金(KJLD14021)资助项目.

摘　　要：针对传统分类模型缺乏对图书数据的多维度语义学习和无法充分学习图书名称、作者和简介等语义信息及其之间联系的问题,提出了一种基于距离注意力的双特征融合的图书分类模型:BTCBLA(BERT-TextCNN-BiLSTM-Attention).该模型采用双通道的方法进行特征提取.针对图书数据包含位置信息的特殊性,在Self-Attention的基础上提出了Distance-Attention,通过加入多尺度相对位置编码解决了在自注意力无法充分理解序列中动态依赖关系的问题;设计了独特的语义融合模块Global-Local-Attention,能够生成权重提取关键信息与语义联系,从而融合局部与全局信息;经过全连接层和Softmax层进行分类,获取多个维度的语义特征,从而具备更丰富的文本语义信息.实验结果表明该模型的准确度达到了91.75%,具有良好的分类性能.Tackling the deficiency of conventional classification models in capturing multidimensional semantic information from book data,which hampers their ability to adequately learn semantic details from book titles,authors,synopses,and their interconnections,the book classification model with distance attention based Bi-feature Fusion(BTCBLA)is proposed,integrating BERT,TextCNN,BiLSTM,and Attention mechanisms.The model employs a dual-channel approach for feature extraction that Channel one,TextCNN,excels in extracting local features from the text,while channel two,combining BiLSTM with Attention,captures global semantic information.Acknowledging the unique characteristic of books data that includes positional information,self-attention is built by introducing Distance-Attention.This innovation,leveraging multi-scale relative position encoding,remedies the self-Attention′s inadequacy in comprehending dynamic dependencies within sequences.Furthermore,the distinctive semantic fusion module is devised,named global local attention,designed to generate weights that emphasize critical information and semantic links,thereby integrating both local and global aspects effectively.The processed features then pass through a fully connected layer followed by a softmax layer for classification.This model harnesses semantic features across multiple dimensions,endowing it with a richer textual semantic understanding.Experimental results affirm its efficacy,achieving an accuracy rate of 91.75%,demonstrating robust performance capabilities.

关 键 词：BERT TextCNN BiLSTM 注意力 图书分类 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]