计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (24): 151-157.DOI: 10.3778/j.issn.1002-8331.2105-0483

• 模式识别与人工智能 • 上一篇    下一篇

基于BG-DATT-CNN网络的方面级别情感分析

余本功,王惠灵,朱晓洁   

  1. 1.合肥工业大学 管理学院,合肥 230009
    2.合肥工业大学 过程优化与智能决策教育部重点实验室,合肥 230009
  • 出版日期:2022-12-15 发布日期:2022-12-15

Aspect-Level Sentiment Analysis Based on BG-DATT-CNN Network

YU Bengong, WANG Huiling, ZHU Xiaojie   

  1. 1.School of Management, Hefei University of Technology, Hefei 230009, China
    2.Key Laboratory of Process Optimization & Intelligent Decision-Making(Hefei University), Ministry of Education, Hefei 230009, China
  • Online:2022-12-15 Published:2022-12-15

摘要: 分析句子针对不同方面的情感极性,深入挖掘评论文本中的信息,为企业生产决策提供建议。针对传统方法多考虑单一层面注意力信息,且基于RNN的模型忽略了局部特征的重要性,而基于CNN的模型不能捕捉长距离依赖的信息的问题,提出了基于双重注意力机制的BG-DATT-CNN模型。在特征表示上,利用BERT对句子和方面词分别进行词向量编码,获得文本的深层语义特征。在特征提取上,设计了双重注意力机制,通过计算两类权重获得综合权重,强化文本的上下文相关特征和方面相关特征。在模型构建上,设计了BG-DATT-CNN网络,结合GRU和CNN各自的优势,Bi-GRU层捕捉文本的上下文全局特征,CNN层包括[K]-Max池化层和TextCNN层,通过两阶段特征提取获取分类的关键信息。在SemEval 2014数据集上的实验表明,与现有的其他模型相比,提出的模型取得了较好的效果。

关键词: 方面级别情感分类, 双重注意力机制, Bi-GRU, [K]-Max池化, TextCNN

Abstract: This paper aims to analyze the emotional polarity of sentences towards different aspects, mine the information in the commentary text deeply, and provide suggestions for enterprises to make production decisions. Traditional methods mostly consider the single level of attention information and the RNN-based model ignores the importance of local features while the CNN-based model cannot capture long-distance dependent information. To solve the above problems, a model named BG-DATT-CNN based on dual attention mechanism is proposed. In terms of feature representation, BERT is used to encode the sentence and aspect respectively, so as to obtain the deep semantic features of the text. In terms of feature extraction, a dual attention mechanism is designed to obtain the comprehensive weight by calculating two kinds of weights to strengthen the context-related features and aspect-related features of sentences. In terms of model design, the BG-DATT-CNN network combines the advantages of GRU and CNN. The Bi-GRU layer captures the global features of context, and the CNN layer, which includes the [K]-Max pooling layer and TextCNN layer, obtains the key information of classification through two-stage feature extraction. Experiments on the SemEval 2014 dataset show that compared with other existing models, the proposed model achieves better results.

Key words: aspect-level sentiment analysis(ALSA), attention machanism, Bi-GRU, [K]-Max pooling, TextCNN