Computer Engineering and Applications ›› 2021, Vol. 57 ›› Issue (8): 119-125.DOI: 10.3778/j.issn.1002-8331.2008-0394

Previous Articles     Next Articles

Research on Stance Detection Based on Dual Attention Mechanism of Heteronuclear Convolution

ZHAO Yuanli, LIANG Zhijian   

  1. College of Big Data, North University of China, Taiyuan 030051, China
  • Online:2021-04-15 Published:2021-04-23

基于异核卷积双注意机制的立场检测研究

赵圆丽,梁志剑   

  1. 中北大学 大数据学院,太原 030051

Abstract:

Aiming at the problem that the target phrase in the current stance detection task appears implicitly in the text, which leads to the poor classification effect, a stance detection model based on Heteronuclear Convolution Double Attention Mechanism(HCDAM) is proposed. This method adopts a three-stage strategy. Firstly, in order to improve the feature representation ability of target phrase and text, the Bert pretraining model is used to obtain the word vector representation with context based on character level. Then, in order to improve the extraction ability of implicit target phrase, the heteronuclear convolution attention way is used to obtain the convolution features with different stance and semantic information. Finally, stance information feature is extracted by using explicit and implicit target phrases through re attention mechanism, and classified by softmax classifier. The experimental results based on NLPCC corpus show that, compared with the Bert-condition-CNN model, the average classification accuracy on the total dataset is improved by 0.108, and the classification accuracy on five topics is improved by 0.146, 0.046, 0.133, 0.047 and 0.056 respectively.

Key words: Chinese microblog, stance detection, attention mechanism, implicit features, deep learning

摘要:

针对当前立场检测任务中目标短语在文本中隐式出现导致分类效果差的问题,提出一种基于异核卷积双注意机制(HCDAM)的立场检测模型。采用三段式策略,为提高目标短语和文本的特征表示能力,采用Bert预训练模型获得基于字符级的包含上下文的词向量表示;为提高隐式目标短语的抽取能力,采取异核卷积注意模式获取含不同位置和语义信息的卷积特征;通过再注意力机制利用显隐式目标短语对文本进行立场信息特征抽取,通过softmax分类器进行分类。基于NLPCC语料的实验结果表明,通过采用异核卷积双注意策略,相比Bert-Condition-CNN模型,在总数据集上平均分类准确率提高了0.108,在5个话题上分类准确率分别提高了0.146、0.046、0.133、0.047、0.056。

关键词: 中文微博, 立场检测, 注意力机制, 隐式特征, 深度学习