文献类型：期刊文献

英文题名：Multi-Target Stance Detection Based on GRU-PWV-CNN Network Model

作者：Li, Wenfa[1,2];Xu, Yilong[3];Wang, Gongming[4,5]

第一作者：Li, Wenfa;李文法

通讯作者：Wang, GM[1];Wang, GM[2]

机构：[1]Inst Sci & Tech Informat China, Beijing, Peoples R China;[2]Beijing Union Univ, Coll Robot, Beijing, Peoples R China;[3]Beijing North Great Wall Photoelect Instruments C, Beijing, Peoples R China;[4]Beijing Tianyuan Network Co Ltd, Beijing, Peoples R China;[5]Inspur Software Grp Co Ltd, Jinan, Peoples R China

第一机构：Inst Sci & Tech Informat China, Beijing, Peoples R China

通讯机构：[1]corresponding author), Beijing Tianyuan Network Co Ltd, Beijing, Peoples R China;[2]corresponding author), Inspur Software Grp Co Ltd, Jinan, Peoples R China.

年份：2021

卷号：22

期号：3

起止页码：593-603

外文期刊名：JOURNAL OF INTERNET TECHNOLOGY

收录：;EI(收录号：20212310464124);Scopus(收录号：2-s2.0-85107205674);WOS:【SCI-EXPANDED(收录号:WOS:000658301700009)】；

基金：This work was supported in part by the National Natural Science Foundation of China under Grants 61972040, and the Premium Funding Project for Academic Human Resources Development in Beijing Union University under Grant BPHR2020AZ03. In addition, we thank Parinaz Sobhani, Diana Inkpen, and Xiaodan Zhu for providing the data used for our experiments, and Ai studio for providing us with a computing platform.

语种：英文

外文关键词：CNN; GRU; Position-weight vector; Multi-target; Stance detection

摘要：To uncover opinions on different people and events from text on the internet, stance detection must be performed, which requires an algorithm to mine stance tags for different targets (people or events). Some text contain multiple targets, and the content describing different targets is related, which results in poor stance detection performances. Therefore, stance detection for such data is defined as multi-target stance detection. To address this issues, a network model composed of a gated recurrent unit, a position weight vector, and a convolutional neural network (GRU-PWV-CNN) is proposed. First, the bidirectional GRU (Bi-GRU) is employed to extract the unstructured features, and a position-weight vector is designed to represent the correlation between every word and the given target. Next, these two forms of information are fused and transmitted to a CNN to complete the secondary extraction of features. Finally, a softmax function is used to carry out the final classification. A multi-target stance detection corpus for the 2016 US election was used to compare the performances of our method and other methods, including the Seq2Seq and AH-LSTM. The experimental results showed that the proposed method achieved well and had a 2.82% improvement in the macro-averaged F1-score.