摘要
[目的/意义]为了解决因微博文本多义性和复杂性导致的谣言检测中语义特征提取不全面的问题,提出了一种结合微博内外多粒度语义的BiLSTM-CNN-ECA模型。[方法/过程]首先,对微博文本从字词句三个粒度级别建模,运用双向长短期记忆网络提取微博内部语义特征,生成事件字向量矩阵和事件词向量矩阵;然后,拼接事件句向量矩阵形成三维文本特征矩阵,输入多尺度卷积神经网络,并行提取微博之间的依赖关系特征;最后,引入高效通道注意力模块赋予通道权重,进行微博谣言检测。[结果/结论]构建的三维文本特征矩阵有机结合了各粒度文本的语义特征贡献,包含更多、更全面的微博语义信息;ECA可有效捕获通道间重要信息,进一步提高了多尺度CNN模型对谣言检测的准确率。
[Purpose/significance]To solve the problem of incomplete semantic feature extraction in rumor detection due to the ambiguity and complexity of Weibo text the paper proposes a BiLSTM-CNN-ECA model combining multi granularity semantics inside and outside Weibo.[Method/process]First of all the micro blog text is modeled from three text granularity levels of characters words and sentences and BiLSTM is used to extract the internal semantic features of microblog and generate event word vector matrix and event word vector matrix;Then the event sentence vector matrix is concatenated to construct a 3D text feature matrix which is used as the input of the three channels of Multi-scale Convolutional Neural Network to extract the dependency relationship features between microblogs.Finally Efficient Channel Attention Module is introduced to generate channel weights and the constructed model is used to detect rumors.[Result/conclusion]The 3D text feature matrix organically combines the contributions of semantic features of different granularity and contains more and more comprehensive microblog semantic information;ECA can effectively capture important information between channels and further improve the precision of multi-scale CNN model for rumor detection.
作者
温廷新
高倩
Wen Tingxin;Gao Qian(School of Business Administration Liaoning Technical University,Huludao Liaoning 125105)
出处
《情报探索》
2023年第5期78-84,共7页
Information Research
基金
辽宁省社会科学规划基金项目“辽宁新型城镇化评价指标体系研究”(项目编号:L14BTJ004)成果之一。
